File access request

N. KUDGUNTI

Compiling dossier…

FILE STATUS: ACTIVE — OPEN TO OFFERS

Nisarg
Kudgunti

AI Engineer who ships multi-agent systems for Fortune 500 clients by day, and built a transformer from raw matrix multiplications by night.

Currently architecting AI vulnerability remediation and RAG systems as an Associate AI Engineer at Techolution in Hyderabad. Before that, an AI Intern at the same company, and a Data Science Intern at AlgoAnalytics. Two years spent in production GenAI. One side project spent rebuilding Llama 3's architecture from scratch, just to know exactly what happens inside it.

Base Hyderabad, India
Role Associate AI Engineer
Experience ~2 yrs production GenAI
Status Open: Pune / BLR / HYD / Remote

· 01 / SUBJECT

About

Nisarg builds the infrastructure layer of generative AI: multi-agent orchestration, retrieval pipelines, and the evaluation loops that keep them honest in production. At Techolution, his work spans vulnerability remediation agents, hybrid RAG systems over thousands of enterprise forms, and natural-language-to-SQL pipelines that turned multi-day reporting into a sub-minute process.

Most of that work means directing large pretrained models. To understand what is actually happening underneath them, he implemented the Llama 3 architecture from first principles in PyTorch: rotary position embeddings, RMSNorm, SwiGLU, the full decoder stack, then pretrained and fine-tuned it himself. That same instinct toward verifiable AI runs through the rest: Counterpoint traces every claim it makes back to a verbatim source quote, and APEX puts a deterministic code-level verifier between its F1 assistant and the user, so an unciteable answer is hedged rather than shipped. All are on file below.

He holds a BE in Computer Engineering with Honors in AI/ML from PVGCOET, Pune, and is Microsoft Certified in Azure AI Fundamentals.

EducationBE, Computer Engineering
Hons. AI/ML, PVGCOET
CGPA8.82
GraduatedMay 2024
CertificationAzure AI Fundamentals (AI-900)
Core stackLangChain, LangGraph,
PyTorch, FastAPI

· 02 / RECORD

Work on file

FILE 03
Feb 2025 to Present
Associate AI Engineer
Techolution · Hyderabad, India · Hybrid
  • Architected a multi-agent vulnerability remediation system for a Fortune 500 media client, integrating Checkmarx PR and push webhooks to detect and fix SAST, SCS, and IaC issues across 50 repositories, with threshold-gated parallelism per vulnerability type.
  • Built a repo-scoped feedback loop on an AlloyDB vector store, embedding accepted, regenerated, and PR-comment feedback with language and type metadata, retrieving the top 5 fixes by cosine similarity to keep outputs aligned to each codebase.
  • Built a hybrid RAG conversational agent over 4,000+ cost estimation forms for an enterprise retail client, using whole-form chunking and LLM-based entity extraction to pre-filter retrieval by known form fields.
  • Designed a key-aware citation system that highlights the exact field and value a generated answer was sourced from, directly inside the original form.
  • Reduced construction schedule reporting from 2 to 3 days down to under a minute with an NL-to-SQL agent over 8,000+ forms, consolidating 80+ tables into 20 UI-mirrored tables to cut hallucination, and a multi-tool pipeline of table selector, SQL writer, validator, executor, and LLM summarizer with table-level schema metadata, DML blocklist validation, cross-project comparisons, trend analysis, and XLSX export.
  • Building a multi-agent platform reducing client project delivery from 3 to 4 weeks down to 3 to 4 days via automated data categorization and segmentation, with gap resolution and impact analysis that identifies low-confidence AI outputs, guides structured resolution, and applies fixes across dependent segments.
FILE 02
Jul 2024 to Jan 2025
AI Intern
Techolution · Onsite
  • Built Project Analyser, a 10+ language GitHub repository analysis pipeline benchmarked at 80k files against the Linux kernel, using Qwen on Cloud Run via vLLM for per-file summarization and Gemini on Vertex AI for executive summaries, service extraction, and dependency graphs.
  • Extended the pipeline with HDBSCAN SOM-based clustering to derive hierarchical feature taxonomies and LLM-generated JSON schemas rendered as ReactFlow architecture diagrams.
  • Reduced VertexAI inference cost by more than 90% by replacing proprietary LLMs with Qwen-2.5-Coder served on Cloud Run via vLLM.
FILE 01
Feb 2024 to Jun 2024
Data Science Intern
AlgoAnalytics · Remote
  • Enhanced a RAG application using LangChain, FAISS, LLM reranking, and innovative chunking strategies to improve RAGAS metrics and Generative AI performance.

· 03 / EXHIBITS

Case files

01 / 03
EXHIBIT A

Llama3-150M-PII-Redactor

A 150M-parameter, Llama 3 style decoder-only transformer, written from scratch in PyTorch rather than fine-tuned through someone else's framework. Pretrained on WikiText-103, then fine-tuned to detect and redact personally identifiable information, so the model that hides the data is the same model that was built to understand it.

The repository keeps pretraining and fine-tuning as clearly separated stages, so the architecture work and the task-specific adaptation can each be read on their own.

Live demo · pattern based, runs in your browser

This is a lightweight rule-based stand-in. The trained model does this with learned attention, not regex. Try the real one on Hugging Face →
ArchitectureDecoder-only transformer, Llama 3 design: RoPE, RMSNorm, SwiGLU feedforward
Parameters~150M
FrameworkPyTorch, written from scratch
Pretraining corpusWikiText-103
Fine-tune taskPII detection and redaction
Fine-tune datasetai4privacy / pii-masking-200k
EXHIBIT B

Counterpoint — Corporate Contradiction Detector

A full-stack AI research tool that ingests a public company's SEC filings, extracts every claim management makes in their exact words, builds a knowledge graph of those claims, and detects contradictions across time. Every finding traces back to a verbatim, verified quote in the source document, never a paraphrase.

The chat assistant scoped to each company is agentic, hybrid, grounding-verified RAG: a LangGraph orchestrator chooses between dense vector search and GraphRAG over Neo4j, is hard-locked to the open company's ticker, and has its output checked against what was actually retrieved before anything is shown.

Live processing runs the full pipeline, fetch through parse, extraction, graph, index, and detection, on any public company on demand, with a bounded free-tier run of a few minutes.

Live app

Open Counterpoint → Pick a company and watch the contradiction graph build. API docs (FastAPI/Swagger) →
PipelineSEC EDGAR filings → paragraph chunks with position anchors → LLM claim extraction with a verbatim quote-span guardrail → Neo4j + Qdrant → contradiction detection
Chat agentLangGraph agentic hybrid RAG (dense + GraphRAG), ticker-scoped, grounding-verified with regeneration fallback
Graph DBNeo4j (Aura) — claims, speakers, documents, topics, contradiction edges with severity and reasoning
Vector DBQdrant (Cloud), FastEmbed BAAI/bge-small-en-v1.5 (ONNX/CPU, no PyTorch)
LLMsOllama Cloud gpt-oss:120b (extraction, judgment, synthesis), gpt-oss:20b (guardrail/intent), config-driven and swappable
BackendPython, FastAPI, Server-Sent Events for token streaming
FrontendReact + Vite, hand-built SVG contradiction graph and citation viewer
Testing & observability63 network-free pytest tests, full LLM call tracing via Langfuse
EXHIBIT C

PayWise — LangGraph Credit Card Optimizer

PayWise answers "which card should I use for this purchase?" with math, not vibes. It runs a 5-stage LangGraph agent that extracts your purchase intent, retrieves reward-rule text from your own cards only, reranks for relevance, computes cashback deterministically in Python, then has an LLM explain the answer in plain English.

Fully traced in Langfuse, scored against a golden eval dataset, deployed via GitHub-triggered CI/CD, all on managed cloud APIs, no GPUs or local models. Deterministic reward math stays completely separated from the two LLM calls, so the one number that matters never hallucinates.

Evals surfaced real issues instead of hiding them: an exhausted LLM API quota that had silently broken prod chat, and a genuine math-node limitation (no rule-specificity tiebreaker), both documented rather than papered over. Also caught a production bug pre-launch, Cloud Run's forwarded-proto header wasn't trusted by uvicorn, silently generating http:// OAuth redirect URIs that would've broken every login.

Live app

Open PayWise → Google OAuth sign-in required. API docs (FastAPI/Swagger) →
ArchitectureLangGraph agent: extract_intent → retrieve_rules → rerank → calculate_math → generate_response, streamed to the UI over SSE
BackendFastAPI, deterministic reward math isolated from LLM calls
FrontendNext.js 14
Vector DBQdrant Cloud
DatabaseNeon Postgres
Embeddings / rerankJina AI
LLMsOllama Cloud / HF Qwen3, swappable
AuthGoogle OAuth
ObservabilityLangfuse, lightweight decorator over pulling in all of LangChain
DeploymentGCP Cloud Run + Cloud Build, path-filtered CI/CD triggers
Eval scorecard15-query golden dataset — Intent extraction 87% · Retrieval 93% · Reward math/winner 67% (all 5 misses trace to one documented root cause)
EXHIBIT D

APEX — F1 Hub

A full-stack Formula 1 statistics app with an AI assistant that cannot make things up. Every factual claim the assistant streams back is traced to a specific database record through an evidence ledger, and a deterministic verifier checks the answer against that ledger in code before a single token reaches the user.

The app itself is a season hub: race calendar down to individual practice sessions, championship tables, driver, team and circuit pages with a WebGL 3D elevation model, live timing, a strategy Pitwall with tyre stints and lap-by-lap replay, and 75 seasons of heritage back to 1950. Data is ingested from Ergast/Jolpica, FastF1 and OpenF1 by a scheduled sync service into MongoDB, so the app serves its own verified data rather than proxying a third-party API at request time.

The design goal was the opposite of a chatbot: the language model narrates retrieved evidence and never derives facts itself. Tools return small pre-joined fact bundles with an evidence id, source and as_of timestamp — relational and arithmetic work happens in Python, so there is nothing left for the model to get subtly wrong. A rules-first router in pure Python sorts questions into fifteen classes across three tiers, so a one-tool question stays a one-tool question and only genuine research fans out to specialist subagents.

Guardrails live in code, not in the system prompt. A verifier node checks that every claim carries a citation, that every cited id exists in the ledger, and that every number in a cited sentence actually appears in that evidence — plus framing contracts, so predictions must hedge and subjective questions never get a verdict. One regeneration is allowed; after that the answer is hedged or stripped. It degrades instead of lying.

Multi-agent was measured, not assumed: a single-agent baseline shipped first, and when one question class came back at 287s and still not converged against the baseline's 51s, that tier was routed back to the flat graph and the result written down.

Live app

Open APEX → Ask the Pitwall Assistant something and watch the activity timeline. API docs (FastAPI/Swagger) →
Agent architectureLangGraph orchestrator + isolated-context specialists (stats-scout, historian, web-researcher, race-analyst); orchestrator holds no raw-data tools so bulky output never reaches the answer window
RoutingRules-first pure-Python classifier, 15 classes across 3 tiers — no model call; cut model calls per answer from ~6 to 2
GroundingEvidence ledger in graph state + deterministic Python verifier: citation presence, id existence, number matching, framing contracts, regenerate-once then hedge
GuardrailsInput guards (scope, injection, PII) run before admission; output guards block zero-tool-call parametric answers. Prompt injection through web search is in the test suite
FrontendNext.js 15 (App Router), React 19, TypeScript, Tailwind, custom glassmorphism system, WebGL 3D circuit models
BackendFastAPI (Python 3.11), Motor async MongoDB driver, Pydantic, SSE token streaming with a live agent/tool activity timeline
AI stackLangGraph, LangChain, deepagents, LangSmith, DeepEval, Tavily, Ollama Cloud inference
DataMongoDB Atlas; Ergast/Jolpica, FastF1 and OpenF1 via a scheduled sync service
Feedback loopLangSmith traces tagged with tier/route/cache/verification; thumbs up-down posts to the feedback API, thumbs-down runs curated into golden-set candidates behind human review; golden set gates CI deterministically
InfrastructureGoogle Cloud Run (4 independent services), Cloud Build CI/CD, Cloud Scheduler, Cloud Storage, Secret Manager, Docker

· 04 / INDEX

Working stack

FW Frameworks
LangChainLangGraphLlamaIndex PyTorchFastAPIFlaskNext.jsReact
AI AI systems
Multi-Agent SystemsRAGLLM Fine-tuning Prompt EngineeringMCPLLM Observability (Langfuse / LangSmith)LLM Evals (DeepEval)
IN Infrastructure
GCPAzureAWS DockerKubernetesvLLMOllama
DB Data
PythonSQLMongoDBRedis FAISSChromaDBPineconeAlloyDB QdrantNeo4jPostgreSQL

· 05 / ADDITIONAL RECORDS

Also on file

Leadership TEDxPVGCOET Licensee and organizer of a 60-member team that hosted 8 distinguished speakers under the theme Punarutthan, the revival of ancient art forms and ideas.
Competition PICT Impetus & Concepts 2024 Third place in the Open Software Domain of the project competition.
Certification Microsoft Certified: Azure AI Fundamentals AI-900, all requirements passed.
Published work The Anthem Of My Heart A poetry collection published in April 2022, his first book.

· 06 / REQUEST ACCESS

Get in touch

Open to AI Engineer, ML Engineer, GenAI Engineer, and Applied Scientist roles, in Pune, Bengaluru, Hyderabad, or remote.