SIDDHESH
KASAT

PYTHON · FASTAPI · AWS STRANDS · BEDROCK · SUPABASE · NEXT.JS

An autonomous agent that runs the business side of freelancing — sourcing, pitching, quoting, invoicing, chasing payment.

Every outbound action queues an approval record carrying the action, its rationale, the messages it cited and the resulting state diff. A separate executor performs the side effect only once a human approves. A virtual clock makes scheduled follow-ups time-travel testable.

A role-based model router defaults to Bedrock and falls back to Groq in development, and a per-run token and cost ledger degrades the model rather than failing the run once the daily cap is hit.

DECISION / NOTHING IRREVERSIBLE HAPPENS WITHOUT A HUMAN SAYING SO.

PYTHON · RAZORPAY API · SQLITE · EVENT SOURCING · GROQ

A revenue-recovery agent for Indian payments that picks the action with the highest expected value.

Measured against two holdout arms on 5,000 payments: ₹5.42L recovered, ₹3.95L of it incremental over the fixed-retry schedule merchants actually run. Compliance sits after the policy as a veto that can only subtract — never outweighed by a large enough number.

Every decision is written to an append-only, hash-chained SQLite ledger with UPDATE and DELETE rejected at trigger level. A 15-point sensitivity sweep confirmed the agent leads both holdouts at every assumption setting.

PROOF / THE LLM CONTRIBUTED −₹238. PUBLISHED THAT RATHER THAN WEAKENING THE BASELINE.

NEXT.JS · TYPESCRIPT · VERCEL AI SDK · SUPABASE · PGVECTOR · COHERE

A conversational job-search agent that harvests live openings straight from company applicant-tracking systems.

Four stages: a rule filter and exact pgvector cosine ranking in one SQL query, job-description gates that drop 37% of the pool, an LLM re-rank scoring true capability overlap, then a link-liveness check that removes same-day closures.

Four LLM tasks routed across Cerebras, Groq and Gemini fallback chains, with per-key rotation, response caching and Postgres-backed usage quotas.

DECISION / RANK INSIDE POSTGRES — 8.35MB DOWN TO 0.64MB PER SEARCH.

NEXT.JS · TYPESCRIPT · LEAFLET · PYTHON · SUPABASE · OPENSTREETMAP

A deployed urban heat-risk platform that scores any map zone 0–100.

Each score decomposes into named drivers — heat exposure, green-cover, cooling-water and shade deficits — ranked by contribution, with the projected score after each proposed intervention.

An interactive Leaflet map with exposure layers, place and street search resolving to the nearest zone, automatic zone naming from nearby OpenStreetMap landmarks, and community heat-hazard reports with optional photo upload.

DECISION / AN EXPLAINABLE ACTION LIST BEATS ONE OPAQUE NUMBER.

NEXT.JS · FASTAPI · CELERY · REDIS · GROQ · SUPABASE · DOCKER

An AI speaking coach: grammar fixes, filler counts, fluency and confidence scores in 10–15 seconds.

The API answers in under 200ms by queueing a Celery task and returning an id, so the work survives a dropped connection. Filler words are counted deterministically from Whisper word segments rather than by asking the model.

Groq Whisper transcription chained into Llama 3.3-70B analysis, returning corrected and native-speaker audio alongside the scores.

DECISION / THE CHEAPEST MODEL CALL IS THE ONE YOU NEVER MAKE.

TANSTACK START · REACT · TYPESCRIPT · GEMINI · GSAP · ZOD

One sentence in, a full scroll-animated landing page out.

Closed vocabulary, open choreography — the model picks from 27 fixed animation intents with clamped parameters and never touches the fragile scroll engine, which a hand-written runtime harness owns identically on every generated page.

Per-section parallel code generation with automatic lint and repair, hand-written deterministic fallbacks, sandboxed iframe previews and single-file HTML export.

DECISION / EVERY GENERATED PAGE STAYS READABLE WITH JAVASCRIPT DISABLED.

PYTHON · LANGGRAPH · E2B · DOCKER · GROQ · STREAMLIT

Natural language becomes code, executed inside a locked-down microVM.

A LangGraph state machine classifies each failure as repairable — syntax, runtime, output-contract — or terminal, where the boundary held and retrying only spends tokens. Termination is enforced structurally by attempt caps and a per-request token, cost and sandbox-time budget.

Untrusted Python runs inside E2B microVMs or a network-disabled Docker container behind a swappable sandbox interface. Benchmarked on 30 real runs at roughly $0.002 per request, with no key material or host file content surviving redaction.

PROOF / 90% FIRST ATTEMPT · 96.7% AFTER REPAIR · $0.002 PER REQUEST.

PYTHON · LANGGRAPH · CREWAI · GEMINI · GROQ

A plain-language business process in, a runnable CrewAI multi-agent project out.

Validated by actually installing and running it — AST import allowlist, fresh-venv install, mocked execution, then domain-specificity scoring — with a repair loop that patches only the failing file and reports honest failure after three bounded attempts.

Generation is constrained to five execution patterns and six real built-in tools — anything else resolves to an explicit mocked stub rather than a hallucinated integration.

PROOF / 134 TESTS ACROSS 21 FILES.