I turn analytics nobody openedinto activation you can measure.

AI judgment analysis for Indian litigators — every claim page-cited.

A production legal-research SaaS for advocates who have to find the ratio decidendi in a 60-page judgment before a hearing. Two-pass LLM extraction fills a 16-field schema, and the prompt forbids inferring a citation: every fact, argument and reasoning point carries a verbatim source quote and page number, or it comes back null.

• 16-field extraction schema where each field carries a verbatim source_quote and page_number — the model is instructed to return null rather than infer a citation.
• Three-tier model router (GPT-4o, GPT-4.1, Claude Sonnet 4.6) behind one interface, with per-model input limits derived from each context window and retry-with-backoff on 429/5xx.
• A legal-transition engine: ~83 hand-curated IPC→BNS, CrPC→BNSS and IEA→BSA mappings across 2,500 lines of seed data, each with delta severity, direction and a practice note.

A voice companion that replies before you notice the pause.

Off-the-shelf voice stacks land a reply four to six seconds after you stop speaking, which is exactly long enough to stop feeling like company. Rezona is a four-surface product — FastAPI backend, iOS app, web app, landing site — built to get an emotionally credible Hinglish-speaking companion talking in under two seconds.

• Cut time-to-first-audio from ~6s to ~1.6s by replacing batch Whisper with a realtime STT WebSocket proxy and streaming TTS sentence-by-sentence — measured across 25 instrumented turns.
• Benchmarked the vendors honestly: ElevenLabs streaming synthesizes in 178ms avg vs Fish Audio’s 415ms, and Fish’s sequential per-sentence WebSocket is ~1.5s slower end-to-end than its own parallel HTTP path.
• Gemini context-caching layer keyed by sha256(model + system prompt), pre-warmed at session start, so one cache entry is shared across every user on a persona stage and memory updates never invalidate it.

AI bedtime stories narrated in a parent’s cloned voice. On both stores.

Bedtime breaks when a parent is travelling or on a night shift. Snoozy generates a fresh ~700-word, five-minute story built around the child’s own name and world, then narrates it in the parent’s own cloned voice. Shipped to the App Store and Play Store across 202 commits in under four months.

• Shipped to both the App Store and Play Store as in.ascentlabs.snoozy (universal iPhone + iPad), including a full Expo 54→56 migration and in-app account deletion to satisfy Apple review.
• Parent voice cloning end to end: multipart upload → Fish Audio model creation → private voice model id, with a DELETE route wired into account deletion so a cloned voice never outlives the account.
• Three interchangeable TTS providers behind one endpoint, each with its own text pipeline — SSML breaks for ElevenLabs, emotion-tag sanitizing for Fish Audio, plain preprocessing for Azure tts-hd.

The whole insurance revenue cycle: lead → policy → commission → invoice.

Indian insurance intermediaries run leads, quotations, policies, renewals, claims and brokerage on spreadsheets and WhatsApp, and nothing reconciles — a policy sells but the commission owed by the insurer is never tracked. This is a production Frappe v15 application that models the full cycle, built over eight phases and 246 commits.

• 34 custom DocTypes (21 parent + 13 child) with status transitions enforced in Python controllers rather than the Workflow engine, so renewal, portability and claim initiation stay auditable.
• A complete commission engine: rate cards resolved per policy year off original_start_date so renewals keep counting from inception, Gross-vs-Net premium bases, instalment proration, flat or percentage rewards, TDS deduction.
• A 1,273-line Policy controller handling Life premium payment terms across five PPT types, including Pay-till-Age derived from proposer date of birth.

An AI cursor that lives in your ERP and points at the field you need.

ERP adoption dies at the form: a new agent opens a 60-field policy screen and has no idea which field matters. A chatbot in the corner cannot fix that — it can describe the Line of Business field but you still have to find it. Clicky injects an animated cursor into every Frappe Desk page that answers the question and then physically flies to the answer.

• Solved LLM UI hallucination structurally instead of by prompt-begging: the frontend gates every POINT target on real DOM visibility, lists collapsed fields as explicitly invalid targets, and voids stale targets on a navigation break.
• A 1,261-line dependency-free vanilla JS overlay: requestAnimationFrame cursor flight, character-streaming speech bubble, hold-to-talk voice input via the Web Speech API, DOM target resolution with fallback diagnostics.
• Harvests live page context per question — route, DocType, visible-vs-collapsed field values grouped by section, child-table row counts, and the snake_case keys of every visible toolbar button.

Creator analytics on the official Meta Graph API, with a Claude monetization engine.

Indian Instagram creators cannot see what their audience is actually asking to buy — native analytics show reach and followers but never connect behaviour to a monetization decision, and third-party scrapers break and cannot see saves or demographics. This is a two-tier SaaS built on the official Graph API instead.

• 21 Meta Graph API functions covering follower/reached/engaged demographics, online-follower hours, media, reels and story insights and paginated comments, behind 22 authenticated REST routes.
• An explicit architectural principle enforced in the schema: fetch fresh from Meta, compute in memory, persist only Claude’s output — the raw_comments table was deliberately dropped in a migration rather than carried as a liability.
• Claude Sonnet 4.5 via AWS Bedrock with markdown-fence stripping, Pydantic response validation and a three-class error hierarchy mapped to precise HTTP responses.

Two-sided agri marketplace with real Razorpay checkout, on iOS and Android.

Mandi traders have no simple way to list stock and sell direct to buyers, and off-the-shelf marketplaces either cost too much to operate or take the seller relationship away. This is a complete marketplace — storefront, seller portal, payments, fulfillment — wrapped into native apps via Capacitor.

• Server-verified Razorpay payments: HMAC-SHA256 checked server-side, an idempotent verify endpoint, and race-safe stock decrement in a single SQL UPDATE — tested against live test mode with a real card and OTP flow.
• Fulfillment modelled on the order *item* rather than the order, so a basket spanning several sellers lets each advance their own items, with forward-only status transitions and cross-seller access control.
• Rotating single-use refresh tokens hashed at rest behind 30-minute access tokens, with the web auth context refreshing proactively so an open tab never hits a surprise 401.

Browser-side sticker studio that cuts you out of a photo and posts to X.

Event badge generators are either a static template with your name dropped in, or a design tool nobody at a hackathon has time to learn. This gives attendees a real canvas, cuts the person out of their own photo entirely in the browser, and then actually attaches the finished card to a post on X — the one thing every share button on the web quietly cannot do.

• A four-rung segmentation fallback (MediaPipe GPU+confidence → CPU+confidence → CPU+category → server matting) where a rung is only marked broken by proof: it threw, or returned an empty mask a lower rung then cut successfully.
• Solved the halo problem by biasing the mask smoothstep ramp to 0.62 instead of 0.5, so the soft edge sits inside the subject — centring it leaves a rim of half-background pixels that reads as a pale outline.
• Real posting rather than a share sheet: OAuth 2.0 + PKCE, media upload and tweet publish against X API v2, with tokens in a single AES-256-GCM-sealed httpOnly cookie and no database at all.

AI-generated video briefings of the Copilot dashboard, for executives.

My Microsoft Global Hackathon 2025 entry, and the only project where I wore both hats officially: I owned the idea, the workflow and the implementation. Executives were not opening the Copilot Analytics dashboard, so the dashboard came to them — as a concise AI-generated video summary of adoption and impact across the org, built from the underlying data and dashboard snapshots.

• Dual-award winner: Most Innovative Project at org level and an Honourable Mention company-wide, out of Microsoft’s global hackathon.
• Wore a dual PM and developer hat — ideation, workflow design and the implementation of podcast generation from data plus dashboard snapshots.
• Framed as a measurement problem, not a content problem: the hack was about quantifying AI’s impact on team performance improvement.

A full product case study, from JTBD to high-fidelity wireframes.

My graduation project for Nextleap’s PM programme, and the work that got me to Top Fellow in Cohort 35. The brief was to increase opt-in for eco-friendly delivery options at Swiggy — a problem where the stated user preference and the actual behaviour point in opposite directions.

• Ran primary and secondary company research, then combined it with prompt engineering to surface user motivations and pain points that surveys alone were missing.
• Defined the Jobs To Be Done and built mind maps to frame hypotheses, then validated them rather than assuming them.
• Landed on a novel, user-centric solution and translated it into high-fidelity wireframes in Whimsical.

IMU-driven electrical stimulation to damp Parkinsonian tremor. Gold medal.

Built for the 7th Inter IIT Tech Meet and won gold. Deep research into Parkinson’s disease led to a lightweight precision glove that senses tremor and actively counteracts it, rather than simply stabilising the load the way weighted aids do.

• Designed a lightweight precision glove to suppress tremor for patients living with Parkinson’s disease.
• Used IMU sensor-driven multichannel Neuromuscular Electrical Stimulation to reduce tremor in the affected joints.
• Gold medal at the 7th Inter IIT Tech Meet, against the other IITs.