EXHIBIT 04-A · ADAPTIVE LEARNING · LIVE
Medha Spark AI
Dynamic roadmaps, an adaptive learning engine, spaced-repetition flashcards and contextual AI explanations. FastAPI · Gemini API · PostgreSQL.
● ARCHIVE OPENFORWARD-DEPLOYED AIAGENTIC SYSTEMSPRODUCTION RAG
Intelligence, with receipts. — after hours.
Karri Vamsi Krishna is a forward-deployed AI engineer — I embed with real workflows and ship LLM systems to production: agentic workflows, RAG over messy data, evals, browser automation. 1.5+ years taking pilots to prod. 1st place, IKDD Agentic AI Challenge / CODS 2025 at IISc Pune. This archive holds seven files. Open them — the room responds to your lamp.
KARRI VAMSI KRISHNA · HYD/IN
SCROLL — THE THREAD FOLLOWS · HOVER — THINGS ANSWERDESCEND 00/07
FILE 01 / 07 · FEATURED EXHIBIT · RAG
The case of the missing pronoun.
Standard RAG chunks a document and severs every “it” from its name. ReferentWeave resolves cross-chunk references before retrieval — without touching the source. Try the interrogation: click the “It” in the document.
CHUNK A · indexedProject Odyssey was officially launched with much fanfare in January 2026.
CHUNK B · indexedIt was delayed by six months due to fuel valve thruster leaks.
QUERY → “Why was Project Odyssey delayed?”
STANDARD RAG → finds Chunk A (name, no answer). Misses Chunk B (answer, no name). Click “It”.
STANDARD RAG → finds Chunk A (name, no answer). Misses Chunk B (answer, no name). Click “It”.
S1
Ingest
Docling / PyPDF
S2
Resolve
Gemini doc-wide JSON
S3
Chunk
1k coupled
S4
Retrieve
Jina + RRF
S5
Rerank
Gemini top-5
Large harness (N=60, 20 docs): R@1 63.3% vs 18.3% · R@3 98.3% · R@5 100%. From tests/eval harnesses in the repo.
LAYER 1 · WHAT +
A reference-aware RAG framework: document-level coreference resolution, dense (Jina v5 1024d, TurboVec 4-bit SIMD) + BM25 (Tantivy) hybrid retrieval, Reciprocal Rank Fusion over top-50, Gemini rerank to top-5, grounded answers with [ChunkID] citations.
LAYER 2 · HOW +
Gemini Flash-Lite resolves each doc in one pass to structured JSON {original_phrase, resolved_entity, confidence}. The enricher appends a [Resolved Context Block] — high-confidence only, exact/normalized match. Source never mutates; generation reads raw text only, block stripped at rerank.
LAYER 3 · WHY IT'S INTERESTING +
In-place pronoun replacement risks hallucinating high-stakes facts. Weaving context alongside the source keeps retrieval smart and generation honest. Parse ~1.5s/page, resolve 2–5s/doc, vector <2ms, ingest <$0.01/100 pages.
LAYER 4 · RESULT +
+26.7pp Recall@1 (N=15), +45pp at scale (N=60). Python + Rust (TurboVec/Maturin), FastAPI. Read the repo ↗
FILE 02 / 07 · MULTI-AGENT SYSTEM
The machine that hunts while you sleep.
JobHunterX — six tool-calling agents in a LangGraph state machine: resume in, tailored applications out, across Greenhouse, Ashby and Lever. Each agent fires as it enters view.
AGENT 1
Profiler
Extracts the candidate profile from uploaded resumes. Structured, verified.
AGENT 2
Planner
Plans targeted search queries from the profile. No spray-and-pray.
AGENT 3
Scout
Searches ATS boards — Greenhouse, Ashby, Lever — with quality gate ≥ 0.60.
AGENT 4
Gate
Zero-token eligibility gate. Kills bad fits before spending a cent.
AGENT 5
Evaluator
Scores every surviving job against the profile.
AGENT 6
Tailor + Browser
Tailors resume PDFs; stealth Playwright agent with human takeover for login, CAPTCHA, MFA — streamed over WebSockets + CDP.
6
AGENTS, ONE STATE MACHINE
85
AUTOMATED TESTS
3
LLM HOUSES · FAILOVER + BUDGET CAPS
HITL
NEVER LOSES TO A CAPTCHA
HOW IT'S WIRED +
Multi-LLM router (Gemini/Gemma · Groq Llama · Mistral) with automatic failover, disk cache, Gemma budget caps and a zero-spend breaker. Search router across TinyFish/Tavily/Exa/Brave. SQLite memory, Jinja2 + xhtml2pdf tailoring. Repo ↗
FILE 03 / 07 · OPEN SOURCE · CLI
The ward where skills go to decay.
Atrophy scans 180 days of your Git history, separates human commits from AI-generated ones, and scores 10 engineering disciplines — so assistance never silently rots your craft. I own the Python/SQLite data layer (~/.atrophy). Presented here in archive tones; the product's own site wears its own neon.
ILLUSTRATIVE RENDERING OF THE REPORT VIEW · DISCIPLINES + DEAD-ZONE RULE (<8 OR >45D GAP) ARE REAL; VALUES SHOWN AS SAMPLE.
SAMPLE · HUMAN VS AI RATIO OVER 6 MONTHS — THE PRODUCT COMPUTES THIS FROM YOUR REPO.
# quarantine the decay. local-first, read-only by default.
pip install atrophy
atrophy scan --days 180
atrophy report # 5-signal AI classifier · Tree-sitter AST > LLM > keywords
atrophy dashboard
atrophy challenge --generate
pip install atrophy
atrophy scan --days 180
atrophy report # 5-signal AI classifier · Tree-sitter AST > LLM > keywords
atrophy dashboard
atrophy challenge --generate
10
DISCIPLINES WATCHED
5
AI-DETECTION SIGNALS
0
BYTES LEAVE (400-CHAR SANITIZED SNIPPET MAX)
THE 5 SIGNALS +
Velocity (lines/min vs your average) · Burstiness (line-length uniformity) · Entropy (char distribution) · Formatter (machine-perfect?) · Message depth (“fix” vs “feat: add JWT…”). Calibrated to your personal baseline — oldest pre-AI commits.
LINKS +
FILE 04 / 07 · TWO LIVE EXHIBITS
Pinned under glass: shipped & clickable.
No mockups. Both deploy to Render — open them in a new tab and interrogate them yourself.
EXHIBIT 04-B · RESEARCH AUTOMATION · ★29
Deep Research Tool
Iterative querying, concurrent multi-engine scraping, Gemini synthesis, resume gap + job-match analysis, automated PDF reports. Zero paid-API-cost scraping.
FILE 05 / 07 · FIELD RECORD · FORWARD-DEPLOYED WORK
Deployed forward, where it had to work in prod.
MAR 2026 → PRESENT
HYDERABAD, IN
HYDERABAD, IN
Junior Data Scientist
@ Sci Tech Patent Art
- Document extraction (OCR + non-OCR, multi-layout): layout parse, Gemini bounding boxes + cropping, LaTeX formulas, line-level coordinates, async transfer with auto-matching → 3–5× lower cost, realtime collaboration.
- Chart construction: replaced an hours-per-document Acrobat grind (~30–40%) with CPU-only OpenCV rectangle-merging + multi-agent spillover detection + citation extraction → 95%+ accuracy, 90–95% lower cost, thousands of pages in minutes.
- Fine-grained layout grounding: per-line coordinates across wild layouts at $0.80 / 1,000 pages.
- Production RAG chatbot: parallel ingestion, leaner orchestration → lower latency + token spend.
- Website-to-PDF + media pipeline: Selenium engine fixing broken Print-to-PDF; keyframe-dedupe + Whisper + LLM summaries at 24/30 FPS.
AUG 2025 → MAR 2026
CHENNAI, IN
CHENNAI, IN
AI Software Engineer L1
@ Gyan Data
- Telemetry AI dashboard — realtime turbine monitoring: autonomous agents, natural-language queries, streamed tokens, live Plotly.
- Graph RAG backend — LightRAG + Neo4j; Frappe ERP customization with Playwright E2E.
- Database-agnostic chatbot — LangChain + LangGraph + Google Auth across OpenRouter/Gemini/OpenAI/Ollama; LangSmith monitoring.
JAN → JUL 2025
CHENNAI, IN
CHENNAI, IN
Data Science Intern
@ Gyan Data
- ERP customization + chatbot development → direct full-time conversion offer.
FILE 06 / 07 · INSTRUMENTS & RECORD
Tools of the trade, schooling in one line each.
AGENTS & ORCHESTRATION
Python · LangChain · LangGraph · RAG · multi-agent systems · tool/function calling · MCP · prompt + context engineering
BACKEND & INFRA
FastAPI · Flask · PostgreSQL · Neo4j · SQLite · Docker · REST · Playwright · Selenium · Git
GENAI & MULTIMODAL
OpenAI · Anthropic · Gemini · OpenRouter · Ollama · Hugging Face · LightRAG · FAISS · Whisper · OpenCV · OCR
ALSO · LANGS · PLAY
SQL · JS · HTML/CSS · C++ · LangSmith · Firebase · Render · Vercel · GCP — English · Telugu · Hindi — chess · competitive programming · photography
| 2025 ★ | 1st Place — IKDD Agentic AI Challenge / CODS 2025, IISc Pune · predictive-maintenance AI. Also: Adobe GenSolve 2nd round · IBC National Hackathon · organized IEEE Mystical Code. |
| 2021–25 | B.Tech, Computer Science — CGPA 8.97 · Gayatri Vidya Parishad College of Engineering (A) |
| 2019–21 / 2019 | Intermediate MPC 97.7% · SSC GPA 9.8 — Sri Chaitanya |
FILE 07 / 07 · CORROBORATION
Don't trust headlines. Read the commits.
Fetching the live ledger from api.github.com/kvcops…
40
PUBLIC REPOS
—
STARS · TOP FOUR FILES
Py
DOMINANT TONGUE: PYTHON
9
FOLLOWERS · EARNED, NOT BOUGHT
CLOSING THE FILE
Have an AI pilot
stuck in notebooks?
Messy docs, flaky pipelines, evals nobody trusts, a demo that dies on real users — I embed with your team and ship it to prod: agents, RAG, automation, proof. One email, real reply.
Vamsikv28@gmail.com ↗