AI/ML Engineer · Generative AI, LLMs, RAG & Agentic AI
ManvikTalwar
I build LLM, RAG and agentic AI systems with measured evaluation, deterministic guardrails and human-in-the-loop safety. B.Tech CSE (AI & ML) at Manipal University Jaipur, three AI and backend internships, and IntelliOps for the Deloitte Capstone Programme.
The Journey So Far
Milestones & Turning Points
Started at Manipal University Jaipur
Began B.Tech in CSE (AI & ML). Built a strong foundation in ML systems, backend engineering, and production-focused AI.
First Deep Learning Project
Started a DWT-CNN invisible watermarking system, which grew into a differentiable encode-attack-decode network reaching 37.3 dB PSNR and 0.97 SSIM.
AI Development Intern at Path Infotech
Deployed "Sofia," an enterprise RAG assistant built on Azure Cognitive Search, with a citation layer that makes every answer auditable.
Interview Mirror
Built a real-time AI interview coach using MediaPipe + Gemini with live feedback, 543 landmark tracking, and low-latency streaming.
SIH 2025 National Semi-Finalist
FloatChat-AI reached the Smart India Hackathon 2025 national semi-finals. Also a Dell x GDG Ideathon finalist and a 4x Student Excellence Award winner.
Agentic RAG for Distributed Content
Built a multi-agent LangGraph RAG pipeline over hybrid vector + BM25 retrieval, with an evaluation harness and citation validation.
Backend Engineering Intern at TechLearn
Shipped the Dashboard Stats API, now the primary analytics endpoint in production, and moved submission retrieval to cursor-based pagination.
AI Engineering Intern at Niva Bupa
Built a claims auto-adjudication engine combining a 7-gate deterministic pipeline with a local LLM for clinical reasoning, with confidence-based routing to human reviewers.
Deloitte Capstone Programme: IntelliOps
Built IntelliOps, an agentic AIOps platform: 75% alert-noise reduction and 41 s median time-to-fix on a live Kubernetes cluster, with governed, reversible remediation.
Featured Work
Projects & Research
From an agentic AIOps platform and local-LLM claims adjudication to agentic RAG and deep-learning research. Each one taught me something the docs never could.
IntelliOps
Agentic AIOps Platform · Deloitte Capstone 2026
A 7-service, event-driven AIOps platform that collapses alert storms into incidents, ranks root causes and runs governed, reversible remediation on Kubernetes. On a live cluster: 75% alert-noise reduction, 41 s median time-to-fix, and autonomy rising from 0% to 76.7% of runs as playbooks earned it from verified outcomes. The LLM agent is confined to a closed 7-verb action vocabulary behind a fail-closed human-approval gate. 833 automated tests.
Claims Auto-Adjudication Engine
Deterministic Rules + Local LLM · Niva Bupa
Health insurance claims engine built during my Niva Bupa internship. A 7-gate pipeline handles policy, member, coverage, waiting-period, exclusion and financial checks deterministically, while a local Gemma LLM on llama.cpp reasons over discharge summaries with GBNF-constrained JSON output. 4-tier confidence routing sends uncertain claims to human review queues.
03Aura
Agentic RAG for Distributed Content
Multi-agent LangGraph pipeline (plan, retrieve, adequacy check, reformulate, synthesize, citation validation) over hybrid vector + BM25 retrieval that abstains instead of hallucinating when evidence is weak. Evaluated with Hit@k, MRR and citation precision: 0.690 Hit@k and 0.549 MRR on local models, blocking every prompt-injection test query.
04Deep Watermarking
DWT-CNN Invisible Image Watermarking
First-authored, end-to-end differentiable encode-attack-decode CNN that embeds a 256-bit payload in the Haar DWT LL subband, with differentiable JPEG in a 6-attack simulator. Trained on 50K images to 37.3 dB PSNR, 0.97 SSIM and 3.2% BER, with 2.4–7x lower BER than a classical DWT baseline. Manuscript in preparation.
05Sofia
Enterprise RAG Assistant · Path Infotech
Enterprise RAG assistant deployed on company servers for internal knowledge retrieval, with a semantic search pipeline built on custom Azure Cognitive Search indexing skillsets for PDFs and a citation layer so every answer links back to its sources.
06FloatChat
Natural Language over Ocean Data · SIH 2025
Smart India Hackathon 2025 national semi-finalist. NL-to-SQL and semantic search over ARGO ocean datasets, replacing hand-written SQL with a conversational RAG interface over 100k+ records.
Aethel
Local-First AI Agent Runtime · In Development
A local desktop agent with permissioned, rollback-safe tools. Reflective Skill Memory keeps or retires learned skills based on their measured success rate.
08Interview Mirror
Real-time AI Interview Coach
Tracks 543 body landmarks at under 100ms latency to analyze posture, gestures, and stress signals. Integrates Gemini, Whisper, and TTS for comprehensive interview feedback.
09MusicGen
AI Music Generation
Generates original music compositions from text prompts using transformer and LSTM-based sequence modeling. Produces structured MIDI outputs with coherent melody, harmony, and rhythm.
Professional Experience
Work & Internships
Niva Bupa Health Insurance
AI Engineering Intern
Built a claims auto-adjudication engine for the ReAssure 3.0 health policy, pairing deterministic rules with a local Gemma LLM on llama.cpp.
- Built a 7-gate pipeline (policy, member, coverage, waiting period, exclusion, financial, state update) mixing deterministic rules with a local LLM
- Forced schema-valid JSON output with GBNF grammar-constrained decoding
- Built 4-tier confidence routing that sends uncertain claims to operations, clinical or senior-auditor review queues instead of forcing an automated verdict
- Built a DAG rule planner (Kahn's topological sort) for audit-replayable plans
- Shipped async FastAPI with per-gate SSE streaming and OCR + LLM extraction from discharge summaries
- Wrote a 67-test pytest suite that runs without a live LLM
TechLearn
Backend Engineering Intern
Built backend analytics APIs for an education platform.
- Shipped the Dashboard Stats API serving real-time KPIs to educators, now the primary analytics endpoint in production
- Replaced offset pagination with cursor-based pagination on submission retrieval, removing deep-page scan cost
- Added composite indexes and restructured trend queries to cut redundant aggregation passes
Path Infotech
AI Development Intern
Deployed an enterprise RAG assistant on company servers for internal knowledge retrieval.
- Built the semantic search pipeline with custom Azure Cognitive Search indexing skillsets for PDFs
- Built a citation layer backed by MySQL tagging so every generated answer rendered clickable source references
- Made outputs auditable for non-technical staff
Recognition
Achievements & Awards
4x Student Excellence Award
Consistently recognized for outstanding academic and project performance.
SIH 2025 National Semi-Finalist
Smart India Hackathon 2025 national semi-finalist with FloatChat-AI: NL-to-SQL and semantic search over ARGO ocean data.
Dell x Google Ideathon Finalist
Finalist for an AI-powered solution at Dell x GDG Ideathon.
Google Big Code Challenge
Ranked top 15,000 across India.
NPTEL Topper
Topped an NPTEL course, the national online learning programme run by the IITs and IISc.
Technical Skills
Tools & Technologies
The stack I reach for when building production AI systems. Always learning, always iterating.