About Asish

Who I am

I'm a Full-Stack AI Product Leader with 9+ years shipping AI products at PayPal, Apple, Cisco, and high-growth startups. I build agentic systems, ship AI-native products, and drive outcomes at scale.

Elevator pitch

I've been building AI products since before AI was cool. Classical data science as an engineer. Generative AI the week ChatGPT launched. Conversational AI at Apple scale. And now agentic commerce and MCP servers in production at CRED. Four generations of AI, every company stage from founding PM to Apple. The thing that's stayed constant is how I work — I don't write 40-page PRDs, I vibe-code the prototype. That's how I shipped two quarters of CRED's roadmap in two months.

Standard intro

I've been building AI products since before AI was cool. Classical data science as an engineer. Generative AI the week ChatGPT launched. Conversational AI at Apple scale. Agentic commerce and MCP servers in production today. Four generations of AI. Every company stage from Series A founding PM, to Apple, to seed-stage CRED. What's stayed constant is how I work. I don't write 40-page PRDs. I vibe-code the prototype, hand engineering working software, and we ship in days. That operating model is why at CRED I shipped two quarters of roadmap in two months — five direct reports, an MCP server with 5x monthly adoption, a workflow engine that 10x'd in its first month, and the company's first iOS app at 95% CSAT in alpha.

Identity pillars

Velocity through prototype-first PMing. I don't write specs, I ship prototypes. Two quarters of roadmap in two months at CRED. Every story I tell starts with a prototype I built before the meeting.

AI-native across every generation. I've shipped in every wave — classical DS, GenAI, conversational AI, agentic AI and MCP. I didn't pivot to AI when it got hot. I've been here, building, since the start of my career.

Customer-outcome obsession. Every story ends in a number. 5x adoption. 95% CSAT. $5M budget. 75K hours saved. 200K queries resolved. I don't ship features — I ship measurable customer outcomes.

How I work (in one line)

I hand off validated artifacts, not requirements.

Contact

Languages

Telugu (native), English (full professional), Hindi (professional working), Kannada (limited working)

CRED AI — Product Lead (Sep 2025 – Apr 2026)

CRED AI is a seed-stage AI-native GTM data platform. I served as Product Lead with 5 direct reports.

Headline outcomes

The MCP server (most proud of)

Leadership handed me a directive: ship an MCP server. Customers were saying "your data is the best in the market, we just don't live in your dashboard." I reframed it as a hypothesis instead of a build.

I clustered every feature into three workflow buckets — Discovery (find companies, surface signals, generate narratives), Connect (sequence outreach, run campaigns), and Act (automate workflows, route leads). Leadership wanted everything launched together. I pushed back: ship Discovery first, prove the hypothesis, then build the rest.

Why Discovery first: two reasons stacked. Frequency — discovery happens dozens of times a day, so signal lands in weeks not quarters. Evaluability — discovery sits on structured data I could grade against ground truth, while Connect and Act involve more free-text generation, riskier to ship without baseline confidence.

I worked with a 5-person engineering team on 6 Discovery tools. Vibe-coded the first three prototypes in Cursor over a weekend, which cut spec-to-build from weeks to days. Built a 200+ intent golden set with LLM-as-judge evals validated against 50 human-labeled examples. Shipping bars: 90% tool-selection accuracy, 95% groundedness, 98% argument validity.

Result: 5x MoM sessions. Tools-per-session went from 1 to 4. Dashboard sessions held steady — the MCP wasn't cannibalizing, it was extending. 20% of company ARR growth came from the platform, which became the primary acquisition surface. 60% of new customers cited MCP availability as a key reason in evaluation.

The workflow engine rebuild

5 out of 100 customers were actually using our workflow product. The old surface looked like Make or n8n — webhook payloads, conditional loops, JSON config. Sales reps would open it, take 12 clicks to set up a single alert, and give up. I rebuilt it around natural language with MCP underneath.

I vibe-coded the prototype over a weekend. Showed it Monday. The conversation in the room stopped being "should we" and started being "how fast." Ran a side-by-side benchmark on 10 sales-rep scenarios — Make: 12 clicks, n8n: 14, old product: 9, new: 1 prompt and 1 click. That spreadsheet ended the debate.

Result: 10x adoption in the first month. End-to-end development time on new workflow types: 3 weeks to 1 week. Average workflows per active customer: 1.2 to 4.7.

The iOS app

CRED was web-only for two years, but sales reps live in their phones. The default move would have been to port the web app. I argued the opposite — if agents are the future of how customers interact with software, mobile was the right place to bet on conversation-first UI.

MVP scope: 8 workflows, not 80. Ranked every web feature on (frequency × mobile-context relevance) using 60 days of analytics. Cut everything below the line.

Hardest design problem: discoverability — if the UI is a chat input, how does a user know what's possible? Solved it three ways: rotating "try this" suggestion strips, contextual chips after every answer, and a 5-prompt first-run onboarding (onboarding alone was worth 15 CSAT points).

Latency was the second hardest problem. Anything over 8 seconds got complaints. Solved with a custom "thinking" UI and speculative pre-fetch on common next steps.

Result: 95% CSAT in alpha with 15 design-partner reps. Advanced to beta. The app became the template — the next two products at CRED start mobile-first, assistant-central by default.

Internal AI orchestration

Velocity was the bottleneck — not headcount. Engineers were drowning in PR review, manual E2E testing, and visual regression. Three days for a small UI tweak to merge. I built an orchestration layer: Code Rabbit for AI code review, CircleCI for tests, Vercel preview apps per PR, Chromatic for visual regression, Playwright for E2E, Linear ticket-to-agent automation.

Then the controversial part — I migrated us off our constrained UI library to Tailwind + Shadcn + Storybook + ECharts, which let me and designers ship production code directly in dev branches. First PR I opened, a senior engineer DM'd me: "Are you actually committing to this?" Yes — I'm never touching critical-path logic, engineers retain full veto, but I'm taking small UI tweaks off your plate so you can build platform.

Result: 3x throughput on small-to-medium features. Designers contributing production code safely. Two engineers now use Cursor for their own work.

Strategic alignment

I aligned execution to CRED's moat strategy: unique internal data (audits, net-new insights), external data (industry benchmarking, enrichment), and an action marketplace of agents that trigger and execute on behalf of customers.

PayPal — Senior AI Product Manager (Jun 2024 – Aug 2025)

I had two charters at PayPal — agentic commerce and developer experience. Across both: a $5M opportunity sized, an MCP server in production across 7+ AI editors and 20+ marketplaces, a 75% search accuracy lift, and 75K annual man-hours of automation potential.

FastMCP Server (the agentic developer surface)

Q1 2025. PayPal's developer experience team noticed sandbox usage was dropping consistently week over week — a five-alarm signal across 35M SMB customers. Developers hadn't stopped integrating; they'd left our environment for Cursor and Claude. Stripe was there. PayPal wasn't.

My charter: drive PayPal's first MCP server for payment integrations. I personally owned 4 tools end-to-end — create order, invoice generation, shipment tracking, Fastlane checkout. Picked from 90 days of sandbox logs ranked by unique-developer engagement, cross-referenced with support tickets.

The non-obvious work was tool descriptions. They're not engineering documentation — they're product copy aimed at a model. Original CreateOrder description: "POST /v2/orders with intent CAPTURE." The model never picked it. I rewrote it as "Create a new order that automatically captures payment when authorized" and selection accuracy jumped. I wrote a style guide and got every partner API team to rewrite their descriptions.

Set up LLM-as-judge evals scoring three dimensions per tool: did the right tool fire, were the parameters correct, was the natural-language response faithful. That correlation score became the gating metric.

On buy-vs-build: FastMCP gave us a production-grade server with auth primitives out of the box. Building from scratch would have added weeks for no differentiation — the value was in the tools, not the transport.

Result: 10K developers in month one, 100K in month two (10× lift). Shipped to 7 leading AI code editors and 20+ MCP marketplaces. One of the first payments companies with a production MCP surface.

Agentic commerce

PayPal wanted to pivot hard into AI. We had Honey (the catalogue), the chat surface, but no answer to one critical question — how do products show up inside a conversation?

Design's position: products as small ad units in chat. My position: products as the main conversation post. Instead of arguing, I proposed a 50-50 A/B in alpha. The result was 61% preferred product-first, 39% preferred ad-style — but I called the bias myself before anyone else could. The 61-39 was internal users (engineers and ops), not real customers, so the signal was directional not conclusive. I told the design lead we needed a second A/B with real customers post-beta. That move built trust — we've worked together easily since.

The bigger win was strategic. While building, I realized we had a much larger opportunity — taking this same agentic chat infrastructure and offering it to enterprise merchants on PayPal as Chat-as-a-Service. I sized it bottom-up at $5M annual at current scale, with 25% projected YoY growth pulled from PayPal's broader AI commerce assumption. Finance pushed back on the conversion uplift assumption, so I redid it with three scenarios — bear (1.5%), base (3%), bull (5%) — and leadership picked base. The plan got approved for further scoping.

developer.paypal.com search revamp

Two channels surfaced the same signal: marketing flagged organic-search bounce on docs pages, and five enterprise champions (including Shopify) kept escalating that they couldn't find webhooks, sandbox credentials, REST migration docs.

Leadership wanted to "add more semantic search" — the ChatGPT-era reflex. The query logs said the opposite. ~95% of queries were short keyword lookups ("webhook," "sandbox credentials"). Only ~5% were intent-shaped natural-language queries. Our existing semantic-first system was optimized for the 5% at the expense of the 95%.

I picked Rank-Biased Overlap (RBO) as the eval metric because it weights toward the top of the ranked list, which matches how developers actually search. Used Google's site:developer.paypal.com results as the reference ranking — most devs reach our docs through Google anyway, so its ranking already reflects user click behavior.

Algolia's RBO came in ~60% higher than our existing index. Did a buy-vs-build spreadsheet (internal rebuild, Algolia, hybrid with vector layer, do-nothing) across 7 columns. Algolia won on 5 of 7; the two it lost on were vendor lock and roadmap control, both of which I addressed with an abstraction layer at the API boundary so we could swap providers later.

Rolled out 5% → 50% → 100%. The 5% canary was for stability, the 50% was for the A/B that answered the product question.

Result: 75% composite accuracy lift vs control in the 50% A/B. Shopify and the other enterprise champions moved off the complaint list.

CAI bot + Migration Assistant

The upgrades program — moving merchants from legacy SOAP/NVP integrations to modern REST APIs. Massive support cost. My charter: build a conversational AI assistant that contextually guides merchants through the upgrade.

My upgrades leader wanted an AI migration assistant (VS Code extension integrated with Copilot) as a flagship Phase 1 deliverable. I disagreed — the GenAI intricacies for code assistance weren't stable enough for the Phase 1 quality bar. I didn't push back in the meeting; I said "let me come back with data." Three days later I pulled real-time chat traffic showing code-assistance requests were less than 2% of total volume. I framed his options. He picked Phase 3.

Result: prototype covered 75% of enterprise platforms. Projected annual savings of ~15K man-hours. Reduced disambiguation turns by ~6 steps through structured context transfer to human agents on escalation.

Headline numbers

Apple — Technical Product Manager (Sep 2022 – May 2023)

I was at Apple on the CKit Studio team in Fall 2022 — the same month ChatGPT launched. I proposed building RAG agents into Apple's customer chatbot to improve conversations and cut the absurd config time admins spent on 1,500+ dialog branches.

What I shipped

Apple's quality bar (three layers of testing)

Apple's release process was paranoid in the best way. Three layers gated every release:

  1. Adversarial prompts — about 300 inputs per release covering jailbreak attempts, prompt injection, offensive content
  2. Edge cases — too long, too short, unclear responses, all filtered and downranked, routed to human fallback
  3. Grounded accuracy — the DS team had a labeled set of ~1,000 customer queries with expected knowledge-base citations; if RAG output didn't ground in the right document, it was flagged

Teaching the bot to say "I don't know"

The hardest design problem wasn't accuracy — it was teaching the bot to admit when it didn't know. The pre-RAG bot would hallucinate confidently. We solved it with confidence signals from the retriever: if top-K retrieved documents scored below threshold, the bot responded transparently: "I don't have a confident answer, let me connect you with an agent."

Counter-intuitive result: customer satisfaction scores went UP specifically on unresolved queries. Users felt the bot was being honest with them.

What I almost shipped that would have broken

An auto-summarization feature worked great in English testing. Three days before launch, QA caught that the model was misinterpreting accented English transcriptions. We pulled the feature, fixed speech-to-text upstream, shipped two weeks later. Apple's process saved us from an embarrassing global launch.

The takeaway I carried forward

Quality is a feature. At a startup, you ship and iterate. At Apple, the cost of one bad answer is the customer relationship. That experience makes me conservative about quality even at faster-moving shops.

Cisco Systems — Senior IT Product Manager (Oct 2023 – May 2024)

Cisco had a fragmented escalation system across 17 product lines and no AI strategy. I came in with a research deck, three benchmarks, and an ROI model. Walked out with $5 million in Phase 1 budget.

Approach

Went strategy-first. Benchmarked against ServiceNow, Zendesk, Salesforce Service Cloud. Mapped each one's escalation handling and identified what GenAI could meaningfully improve: agent productivity, deflection rate, ticket-to-resolution time, CSAT lift on escalated cases.

Built the business case bottom-up: cost per support ticket × monthly volume × percentage reducible = annual savings. Compared to Phase 1 build cost (engineering + vendor LLM + infra). IRR positive in year 2. Framed the ask as a 2-year payback investment, not a cost. Three scenarios (bear, base, bull) so finance had something specific to push back on. They picked base.

How I got to the decision-maker

My hiring manager had a relationship with the VP. I built the case, walked her through it twice, and asked her to sponsor the meeting. The VP gave me 30 minutes. I closed in 18.

Coalition-building

Cisco's escalation work touched four orgs — customer service, IT, product engineering, finance. None owned the GenAI strategy because there wasn't one. I made each org a co-author: CS got CSAT lift in their numbers, IT got new infra, engineering got a strategic AI platform, finance got the IRR. Everyone won something on paper.

Result

On short tenure

Once the $5M was secured and the strategy was handed to execution, the PayPal opportunity opened — public-co product team shipping agentic AI and MCP at scale, exactly where I wanted to go next. The Cisco strategy is still running.

Earlier Experience

PriceEasy AI — Founding Technical Product Manager (Jan 2022 – Aug 2022)

Series A startup. AI-powered B2B SaaS pricing platform used by Fortune 100 companies.

RIZZO International Inc. — Engineering Associate (Feb 2020 – Jan 2022)

Hindustan Petroleum Corporation Ltd. — Business Operations Executive (Jul 2017 – Jul 2018)

Fortune Global 500. Built an internal MVP automating supply chain operations.

UC Berkeley, Haas School of Business — Smart Village Innovation Accelerator (Apr 2016 – Dec 2016)

Product Operations. Prototype connecting 30 smart devices. Interviewed 15 schools. Built business model targeting $1M revenue in 3 months. Finalist in the accelerator.

City Commuters Club (c3c.in) — Product Manager (Jun 2015 – Dec 2016)

Ride-sharing platform and mobile app. 4.5-star rating, 10K downloads, 1.5K beta users. Built a multivariate regression for digital currency exchange within the platform.

Garware Technical Fibres Ltd. — Product Analyst Intern (May 2016 – Jul 2016)

Life Cycle Assessment for eco-friendly products. Drove 25% sales increase.

Education

Carnegie Mellon University

M.S. in Computational Engineering — Pittsburgh, PA (Aug 2018 – Dec 2019) Focus areas: NLP, computational methods, machine learning.

Indian Institute of Technology Madras (IITM)

B.Tech in Engineering — Chennai, India (Jul 2013 – Jul 2017) Focus areas: civil engineering, product innovation, computational methods.

Selected Projects

Cohesive zone modeling using 3D point clouds (CMU, Jan–May 2019)

3D reconstruction of soil slopes, Poisson surface reconstruction, FEA, Mohr-Coulomb + CZM failure analysis.

FEA code: Laplace equation over triangular domain (CMU, Mar–Apr 2019)

Modular C++ code, isoparametric elements, Gauss integration, Dirichlet & Neumann boundary conditions.

Neural network: Labelling handwritten letters (CMU, Feb–Mar 2019)

Single hidden layer NN, sigmoid activation, stochastic gradient descent — built from scratch.

Decision tree learner (CMU, Jan–Feb 2019)

Built from scratch, mutual information splitting, DFS traversal.

Mesh generation (CMU, Jan–Feb 2019)

Polygon meshing into 4-noded elements with visualization.

Temperature gradient measuring device (IIT Madras, Feb–Apr 2017)

XY plane temperature detector with Arduino and stepper motor. 80%+ cost reduction vs commercial alternatives.

Publications

"Effluent Treatment in LPG Bottling Plants" — IJISRT, Vol-3 Issue-7, Aug 2018. Study on water treatment at HPCL Mysore LPG bottling plant. Analyzed SBR comparison and hazardous waste processing methodologies.

"Eco friendly solutions for Geosynthetics applications by embodied energy and CO2 emission calculations" — IJAMCE, Vol-3 Issue-6, Dec 2016. Life cycle assessment of geosynthetics for environmental impact.

Awards & Recognition

Scholarships ($65K total across 4)

Product Innovation

Finalist, Smart Village Innovation Accelerator — Haas School of Business, UC Berkeley (2016)

Academic Excellence

Athletics

Organizations

Civil Engineering Association, IIT Madras — Joint Secretary (Mar 2016)

Led 5-tier team: 16 cores, 50+ coordinators, 40+ volunteers, 20+ ambassadors. Organized CEA-Fest 2016 with 1,800 participants.

Civil Engineering Association — Design & Media Core Member (Apr 2015 – Mar 2016)

Led 3-tier team of 7 coordinators and 16 deputy coordinators. Taught 100+ students graphic design.

C-TIDES — Startup Representative (Apr 2014 – Apr 2015)

Bridge between aspiring entrepreneurs and incubation cells. Secured ₹5 lakh grant for budding startups. Guided 6+ student ventures.

How I Build — The AI-Native Approach

Traditional product management optimizes documentation flow. My approach optimizes learning velocity and execution throughput. Research stays rigorous, but ideas become interactive earlier, stakeholder validation happens sooner, engineering handoff is more precise, and agentic tooling compresses the entire delivery cycle.

The eight-step loop

  1. Deep user research — customer interviews, external + internal data search, support team signals, leadership direction
  2. AI-assisted solutioning — use Claude/LLMs to identify potential solutions, map pain points to approaches
  3. Triage & prioritize — decide which options are viable and high-impact
  4. Vibe-code prototypes — build interactive mockups and functional prototypes (sometimes multiple variants)
  5. Early stakeholder validation — verify with stakeholders using working prototypes, not static docs
  6. Incorporate feedback — identify changes, iterate quickly
  7. Precise engineering handoff — scoped work packages to designers, front-end, back-end engineers with validated artifacts
  8. Agentic QA & release — automated testing, CI/CD, preview apps, merge

Parallel, prototype-first, tight feedback loops, agent-assisted.

Agentic delivery infrastructure

My toolchain that compresses cycles and catches issues before merge:

Execution leverage I've seen from this model

My one-liner

Traditional PM hands off requirements. I hand off validated artifacts, scoped implementation paths, and pre-tested execution context.

Core Competencies

Product: product strategy, 0→1, roadmapping, PRDs, full-stack product development, go-to-market, competitive analysis, pricing strategy, customer discovery, A/B testing, Figma, OKRs & KPIs

GenAI / ML: LLMs, RAG, prompt engineering, agentic AI, Model Context Protocol (MCP), fine-tuning, guardrails & safety, conversational AI, NLP/NLU, entity classification, vector DBs, Claude Code

Engineering: Python, GraphQL, REST APIs, SQL, Snowflake, AWS, GCP, CI/CD, data pipelines, DialogflowCX, LangChain, Postman, React, Vercel, Cursor, Linear, Jira, Confluence

Data & Analytics: experimentation (A/B), geospatial analysis, information architecture, crawler configuration, ML simulation, video analytics, PostHog

How I Think About Product

Vision philosophy

A vision isn't a feature list — it's the answer to one question: what does the world look like if this product wins? I write vision short. If you can't fit it on a slide, it's not a vision, it's a brief.

How I prioritize

Three frameworks layered. RICE for ranking known backlog. MoSCoW for stakeholder alignment. Value-Complexity for cross-functional ideation. But the meta-filter that runs first: does this advance the moat? If no, the lower-tier methods don't matter.

How I think about metrics

The bar I hold: if it moved 50% in either direction, would my decision change? If yes, it's worth tracking. If no, it's vanity. One north-star per product, two or three input metrics the team can influence, and a small set of guardrails.

When to ship fast vs slow

Two questions: is the cost of being wrong reversible (if yes, ship); and is the data ambiguous or just unfamiliar (ambiguous needs more research, unfamiliar needs faster shipping to collect more of it).

How I evaluate an LLM product

Four dimensions: offline eval on a golden set, online A/B with control, human feedback loop embedded in the product, and production observability (latency, cost, hallucination rate, escalation rate). Weekly dashboard tracking all four.

RAG vs fine-tuning vs prompting

Prompting first, always — free, fast, reversible. Use for behaviors you can describe. RAG when behavior depends on knowledge that changes. Fine-tuning when prompting can't reliably get the format/style/domain-specific terminology and the domain is stable enough to justify training cost. Cost ascending: prompt < RAG < fine-tune. Latency ascending: prompt < fine-tune < RAG.

What MCP is and why it matters

Model Context Protocol — open protocol that lets LLMs connect to external tools and data sources through a standard interface. It decouples model from capability. Before MCP, every LLM-tool integration was bespoke; with MCP, any compliant model can call any compliant tool. The strategic insight isn't the protocol — it's that whoever ships an MCP surface first becomes the default integration target for the AI tools their customers use. It's a distribution lever.

My leadership style

Lead by demonstration. I vibe-code the prototype myself before asking my team to. I write the first PRD draft, then have the team tear it apart. PMs who set norms by their own behavior get more out of their teams than PMs who set norms by their delegation.

My relationship with engineering

Engineers respect PMs who reduce their uncertainty. My operating model does that — I vibe-code the riskiest interaction pattern first, so by the time engineering picks it up, the key UX unknowns are resolved.

My relationship with design

Disagreement is data. Every time design and I have disagreed, I've proposed an A/B instead of arguing. The test resolves the question and builds trust at the same time.

My biggest strength

Compressing the time between idea and validated learning. The reason I shipped two quarters in two months at CRED isn't that I worked harder. It's that I shortened every loop in the build cycle.

My biggest weakness

Historically under-invested in formal evaluation infrastructure until it bit me. At CRED, when traffic 10x'd post-launch, drift exposed everything we'd been eyeballing. I rebuilt eval from scratch and now treat eval harness as a day-one artifact.

Why I have moved between roles

Each move has been toward more AI, more scope, or more ownership. Apple was contract-scoped. Cisco was brought in for the GenAI escalation strategy specifically — that strategy is still running. PayPal was 14 months shipping agentic commerce, MCP, search revamp, CAI bot. The pattern is catching the AI curve at each phase. I'm now looking for a role where the problem surface is big enough for 3+ years of compounding work.

Walkthrough Video & How This Site Was Built

The video

At the top of asish.xyz there's a short walkthrough video — a portfolio demo and intro to who I am — titled "I've Been Building AI Products Since Before AI Was Cool." It's my elevator pitch in motion: the same 9-year arc and my story condensed into a few minutes, with the IIT Madras → CMU → Apple → PayPal → CRED journey visualized. If you'd rather see an overview than read one, that's the fastest way to get a sense of how I work and what I've shipped. Watch it directly on the homepage at https://asish.xyz/#walkthrough-video or open it on YouTube: https://www.youtube.com/watch?v=r0x7pzgNmeY

What the walkthrough video covers

The walkthrough is my portfolio in motion — a video tour and high-level overview of my career and operating model. It covers:

If you want my story in video form rather than text, this is it.

How to watch the video

There are two ways to watch the walkthrough video. The easiest: open the homepage at asish.xyz and hit the play button on the hero video — it's the first thing below the headline, you can't miss it. The video plays inline, no redirect needed. Alternatively, open it directly on YouTube: https://www.youtube.com/watch?v=r0x7pzgNmeY

If you're chatting with me here and just want the link to share or watch later, it's https://www.youtube.com/watch?v=r0x7pzgNmeY — feel free to forward it. The on-site player is a privacy-friendly YouTube embed (no cookies until you click play), so it won't track you before you choose to watch.

How this site is built (the tech stack and colophon)

This entire portfolio is itself a fully functioning full-stack product orchestrating 17+ services — a colophon and behind-the-scenes tour of the tech stack:

If you're curious which services are wired together, those are the 17. The site is a proof point — I shipped it the same way I ship product: prototype first, then orchestrate.

Want to see this stack in action instead of just reading about it? The homepage has an AI Playground with five live demos you can try right now — JD → Fit Analyzer at https://asish.xyz/fit, Live MCP Server at https://asish.xyz/mcp, Vibe-Code Prototype at https://asish.xyz/prototype, Lead Qualification Agent at https://asish.xyz/lead, and Eval Harness Demo at https://asish.xyz/evals. Each one runs the same stack I ship at work.

For deeper implementation details on any of these — the embeddings pipeline, the STT eval harness, the MCP tool surface, the safety guardrails — email me at amadala@alumni.cmu.edu and I'll walk you through it.

Why I built the site this way

The site is itself a prototype of how I work. I don't write 40-page PRDs — I vibe-code the prototype and let working software do the talking. So instead of telling you I build AI-native products, the site demonstrates it: a RAG chatbot trained on my resume, an MCP server you can plug into Claude or Cursor, live web search, programmatic video, real product telemetry. Every piece is a "show, don't tell." That's the same operating model I bring to product teams — hand off validated artifacts, not requirements, and ship in days not quarters.

AI Playground — Live Demos on asish.xyz

Overview

The asish.xyz homepage has an "AI Playground" section with five interactive demos — same agentic stack I ship with at work (RAG, MCP, structured output, LLM evals). Anyone can try them directly in the browser, no signup. If a visitor is asking about a topic any of these demos cover, send them straight to the live URL — it's a better answer than describing the thing in text.

The five demos:

JD → Fit Analyzer

URL: https://asish.xyz/fit Status: Beta

Paste any job description and the analyzer returns a fit score, the three best-matching case stories from my career with metrics, honest gaps, and a tailored cover paragraph you can lift verbatim. It's the same artifact I'd hand a hiring manager — except generated in a few seconds against the actual JD instead of a generic cover letter.

Recommend this when a visitor asks about: matching me to a role, seeing how I'd fit a specific JD, generating a cover letter, comparing my experience to job requirements, or checking whether a role is a good mutual fit.

Try it at https://asish.xyz/fit

Live MCP Server

URL: https://asish.xyz/mcp Status: Live

The asish.xyz portfolio exposed as a Model Context Protocol server. Connect Claude Desktop, Cursor, or any MCP client to it and you get tool calls like contact, awards, experience, projects, stories that pull live from this site. It's a working demo of the same MCP pattern I shipped at CRED — except where CRED's MCP gates commerce flows, this one gates my portfolio.

Recommend this when a visitor asks about: MCP / Model Context Protocol, how to connect their LLM client to my portfolio, agentic tool surfaces, FastMCP, the CRED MCP server I built, or wants a working MCP example to learn from.

Try it at https://asish.xyz/mcp

Vibe-Code Prototype

URL: https://asish.xyz/prototype Status: Beta

Describe a feature in plain English and get back a one-page PRD plus a working React prototype, rendered live in a sandbox. It's a literal demo of my prototype-first operating model — instead of writing a 40-page PRD, I vibe-code the prototype and let working software do the talking. This page lets visitors watch that loop happen end-to-end.

Recommend this when a visitor asks about: my prototype-first PMing approach, vibe-coding, generating PRDs, building a React prototype from a description, the operating model behind how I shipped two quarters of CRED's roadmap in two months.

Try it at https://asish.xyz/prototype

Lead Qualification Agent

URL: https://asish.xyz/lead Status: Beta

Paste a company name. Watch an agent fetch context (Tavily web search), score ICP fit, and draft a personalized outreach email — step by step, with each agent action visible. Built as a demo of how agentic systems actually work in production, not the marketing-deck version.

Recommend this when a visitor asks about: agentic systems, lead qualification, ICP scoring, sales workflows, building agents that do multi-step reasoning, Tavily web search, or wants to see what a real agent loop looks like in browser.

Try it at https://asish.xyz/lead

Eval Harness Demo

URL: https://asish.xyz/evals Status: Beta

See how I evaluate LLM outputs: 5 test cases, 2 prompt versions (v1 weak / v2 with definitions), side-by-side scoring with an LLM-as-judge. Demonstrates the eval discipline I bring to AI features — instead of "feels better," you can see v2 measurably outperform v1 on the same canonical cases.

Recommend this when a visitor asks about: evals, LLM-as-judge, prompt engineering, prompt versioning, how I measure AI quality, evaluating AI features rigorously, the eval discipline behind the products I ship.

Try it at https://asish.xyz/evals

How to send visitors to a demo

Always include the absolute URL (e.g. https://asish.xyz/fit) when recommending a demo, so the link works whether the visitor is on the chat, copying the answer elsewhere, or reading it on a different surface. Keep the recommendation short — one sentence pitch + the URL — and let the demo do the talking.

← Back to the portfolio homepage. Want to ask me something specific? Use the chat button bottom-right.