PavanD.
Deep learning × VLSI — I teach machines to find the fault in silicon. Bangalore.
I build ML systems that find faults in silicon — plus LLM and deep-learning tools — and tell you honestly when they don't work.
Chips fail in ways you can't see. My work lives where deep learning meets VLSI: reading how a circuit misbehaves and pointing at the exact place it broke. My flagship, Faultiva, does that with a hybrid fault-intelligence stack — and every project since has been a harder test of the same idea.
I care about the unglamorous part: reproducible runs, provenance on every result, and reporting the outcome that actually happened — including the ones that don't beat the baseline. A result you can't reproduce, or an overclaim, is worth nothing in hardware.
Beyond silicon I work across the ML stack: deep learning in PyTorch, LLM agents with typed specs and guardrails, local inference with llama.cpp, vector search, and the full-stack and Windows tooling to ship it all. I presented a paper on an online voting system at ICEST 2026, built an AI resource-allocation tool in a 36-hour hackathon, and I work in English, Kannada and Hindi.
- Based
- Bangalore, India
- Focus
- Deep learning × VLSI · LLM systems
- Studying
- Electronics & Communication · final year · P.A. College of Engineering
- Building now
- TinySPICE
- Open to
- ML / AI-research internships · Jan 2027
- Languages
- English · Kannada · Hindi
- Method
- Reproducible · honest · shipped
- Ask me
Selected work
-
Find the broken gate in a chip from its output alone. Detection, localization and structural verification, as a Windows app or a 4.7 MB Python package. Hand it what a circuit produced and what it should have produced: it says whether the chip is faulty and which gates could be responsible. Stuck-at fault campaigns across four cores (AES, SHA-256, OpenTitan HMAC, PicoRV32) with Yosys and Verilator, a 3-stage pipeline, and a PyTorch verifier over netlist topology. It also states plainly that only 65.7% of faults are uniquely localizable. Undergraduate major project, with a Zenodo DOI.
- ~0.5 msper query, CPU-only
- 0false alarms over 7.85M transactions
- 0.9166verification ROC-AUC
- 36,784 → fewcandidate fault sites
PythonVerilatorYosysstuck-at faults93,185-param verifierWindows installerDOISource ↗ Download v1.1.0 ↗ DOI ↗ -
A graph-aware hybrid model for OpenTitan HMAC faults. A research prototype that predicts whether a stuck-at fault will be detected, from the circuit's graph structure. It beat conventional and plain deep-learning baselines, then lost more performance than allowed on separate validation vectors. Both results are published side by side: an honest baseline, not a highlight reel.
- 22,839fault sites
- 45,678fault instances
- 47,617graph edges
PythonPyTorchgraph featuresOpenTitan HMAC-SHA256SA0 / SA1Source ↗ -
I spent two days optimising the wrong metric. A localization study whose result is a negative one, proved rather than shrugged off. Six formulations falsified, one impossibility proof, a structural bound. Grow the candidate catalogue from one circuit to four and exact-site accuracy collapses 19.2×. No model can fix that. The acceptance bar is marked NOT MET, in red, on the front page.
- 19.2×accuracy collapse (0.549 → 0.029)
- 6formulations falsified
- 76stages · 7.85M transactions
Pythonpre-registered contractsaudit trailnegative resultsSource ↗ -
Fault localization that survives noise. Chip behaviour becomes a vector; the nearest known fault behaviours locate the fault. Exact dictionary lookup dies at a single flipped bit while approximate vector search keeps working.
- 366/366c432 faults localized
- 833/833c880 faults localized
- 96–99%recall at 32 flipped bits
- 0%exact lookup at 1 flip
PythonEndee vector DBDockerISCAS-85stuck-at simulationSource ↗ -
An LLM agent you can't talk into lying. Ask for a MOSFET characterization in plain English. The LLM only parses the request into a typed spec; deterministic code validates it, plans inside hard budgets, runs real ngspice on SKY130, extracts the measurements, recovers from named failures and stores a hash-verified, replayable record. On TinySPICEBench (24 pre-registered tasks), a direct LLM-to-tool pipeline claimed success on runs that had failed 30 times out of 72; the full system did it 0 times. Negative results are reported, not hidden.
- 30/72 → 0/72false successes, direct vs full
- 0.50 → 0.86task completion
- p = 0.001sign test
- 269tests · 18/18 reproduction checks
PythonngspiceSKY130llama.cpp (Qwen3.5)SQLite provenancepre-registrationSource ↗ -
Can adaptive sampling beat brute force? Can adaptive sampling train a useful drain-current predictor from fewer SPICE simulations than uniform sampling? ngspice data on a pinned SKY130 revision, provenance manifests on every curve, a PyTorch MLP on nested budgets (32 to 256 curves) against MC-dropout active learning at equal budget, and a width-held-out test against a tree baseline. Active learning led at 64 and 128 curves and lost at 256: reported as mixed.
- 64 · 128budgets where active learning led
- 256where it lost, reported
- 45public NMOS files benchmarked
PythonPyTorchngspiceSKY130 NMOSactive learningSource ↗ -
Project memory for developers, with receipts. Scans a codebase read-only into memories with file, line-range and content-hash provenance, flags the memories a code change may have made wrong, keeps contradictions open, and answers with citations or an honest “evidence not found”. Serves coding agents over a read-only MCP server. No LLM in the loop. Shipped as an installable Windows app, with a security audit that took dependency advisories from 93 to 0.
- 0.80 → 0.98Recall@5, keyword vs semantic
- 83%useful abstention
- 572backend tests · CI green
- 93 → 0audited advisories
FastAPIReact/TypeScriptSQLite FTS5MCPPyInstallerInno SetupSource ↗ -
A tiny transparent audio visualiser for Windows. Whatever your PC is playing becomes a quiet wisp of light floating on your desktop. It taps the final audio mix through WASAPI loopback, so nothing needs configuring per app. 11 animations, auto-gain and beat detection, starts with Windows. No installer, no runtime, no dependencies.
- ~180 KBsingle exe
- ~4 MBRAM
- 11animations
C++Win32WASAPIFFTw64devkitSource ↗ -
One face, one vote. A voting application with user authentication, a one-vote-per-user integrity check that blocks duplicate votes, face verification before the ballot opens, and live result tallying.
- 1vote per verified face
PythonFlaskDeepFaceauthenticationSource ↗ -
An AI resource-allocation tool for emergency teams. Built at Nexora Hackathon 2025 in Bengaluru, in a four-person team, taken from concept to a working demo inside the time limit. Matches scarce supplies and teams to the places that need them most.
- 4person team
- Nexora 2025Bengaluru hackathon
PythonAIresource allocationhackathonSource ↗
Research and publications
- 2026FaultivaFind the broken gate in a chip from its output alone. Archived and citable.Software · Zenodo DOI
- 2026CircuitSage V2 — the audit trailSix falsified formulations, one impossibility proof and a structural bound, with the contracts frozen before measuring.Research repository
- 2026CircuitSage HMAC — frozen v1Graph-aware SA0/SA1 detectability on OpenTitan HMAC-SHA256, with the validation drop published.Research repository
- 2026TinySPICE & TinySPICEBench24 pre-registered tasks: 30/72 false successes for a direct LLM pipeline against 0/72 for the full system.Research repository
- 2026IVForgeAdaptive sampling against a uniform baseline for transistor I–V characterization, mixed result reported.Research repository
Certificates
Hover a name to preview · click to open the full certificate
Toolkit
Silicon
- ngspice circuit simulation
- SKY130 PDK open 130nm
- Yosys · Verilator Verilog / RTL
- ISCAS-85 benchmark circuits
- Stuck-at faults simulation & dictionaries
- I–V characterization MOSFET
Machine
- Python daily driver
- PyTorch · scikit-learn deep & classical ML
- LLM agents typed specs · guardrails
- llama.cpp local inference
- Vector search ANN retrieval
- Active learning surrogate models
Systems
- FastAPI · Flask · React full-stack apps
- SQLite · FTS5 local-first data
- C++ · Win32 native Windows tools
- MCP agent tooling
- PyInstaller · Inno Setup installers
Cloud & DevOps
- AWS cloud
- Azure cloud
- Vercel web hosting
- Cloudflare edge · Pages
- Docker · WSL2 reproducible envs
- GitHub Actions · Git CI & shipping
- Linux daily shell
Let's find
the fault.
I'm looking for ML / AI-research and silicon-adjacent internships starting January 2027, Bangalore or remote. If you have a hard problem and want someone who ships and reports the truth, write to me.