ML systems engineer · Founder at ScholarLM
I make ML and developer infrastructure fast, and I prove it with numbers: Metal GPU kernels, Rust and Go runtimes,
and code-intelligence tools for AI agents, built as modules that every new project reuses.
Open to opportunities · bharath@vbcr.dev
- Now: founder and sole engineer of ScholarLM, an AI research platform (React 19, Rust, Go and Python) that writes fully-cited, grounded manuscripts (97.7% [
85/87] precision on public LitQA2 without a supplied paper corpus). In parallel I build the Apple-silicon GPU, Go↔Rust contract, and code-intelligence stack below. - Strongest results: an Apple-silicon Metal 4 multimodal runtime (3.06–4.37× faster in bf16 than PyTorch MPS while fitting 4K–6K ragged workloads in ≤1,911 MiB where PyTorch OOMs at 33.6 GiB), LoRA activation checkpointing and weight-gradient skipping cutting Metal device peaks by −78% and −44%, a code-graph indexer 1.6–21× faster than four other tools on every corpus tested, GitPulse cutting 30k-commit graph refresh to 1.11 ms (
25.19 → 1.11 ms) and watcher wakeups from 117.3 → 2.7/s, and a Go↔Rust call made 8.5× cheaper. - Before: Associate Researcher at the Adidas Center for Engagement Science (ASU, 2023–2025), and stem-cell research at Texas Tech University Health Sciences Center.
- Education & papers: M.S. Biomedical Engineering, Arizona State University (2024) · 2 peer-reviewed papers (2025).
- Upstream: Initiated the official Eclipse
iceoryx2Go bindings (eclipse-iceoryx/iceoryx2-go: proposal, repo provisioning, first PR under review); merged patches inrust-lang/libc(0.2.191) andobjc2; Apple Silicon MPSNaNdiagnosis & reproducer indecider-ai(1.8.2).
- Eclipse iceoryx2 — proposed and scaffolded the official Go bindings in
eclipse-iceoryx/iceoryx2-go(#225, #2104). Provisioned the repository and branch rulesets with the project leads (.eclipsefdn#28) and built the Step 1NodeBuilder→Nodecgopackage over theiceoryx2-ffiC API with a release-pinnediceoryx2submodule (iox2_*_storage_tABI parity),golangci-lintv2, and Go 1.26/1.27 Linux/macOS-raceCI, now open for review as PR #1 (first validated on the fork). Committer election follows a review track record; the binding is not published yet. - rust-lang/libc — added Apple
F_SETNOSIGPIPE(73) andF_GETNOSIGPIPE(74) (#5609), shipped inlibc 0.2.191. Upstreamed from completion-pipeSIGPIPEsuppression in Gusset. - objc2 — marked
MTLTensorDescriptorconsumersunsafeinobjc2-metal(#870, #782, found while building Metal 4 tensor ops in tessl): Metal segfaults on a nil stride or an oversized tensor instead of returning an error. Also restricted the AppKit examples to macOS (#871). - Mapika/decider (
decider-ai) — diagnosednon_blocking=Truehost-to-MPS copies of temporary CPU tensors returningNaNinscaled_softmaxandread_slots(#21). The reproducer was adopted upstream as a temperature regression test indecider-ai 1.8.2(fix by the maintainer). Verified on M5 Pro:1,697/1,698choice parity with CPUfloat32. - Electrobun — reported Windows packaging resolving
rceditfrom a baked CI path (#298) and submitted a fix (#299). That pull request was closed without merging. Also diagnosed the hardcoded Windows icon path (#235).
Each figure comes from the project's own benchmark record, with its conditions. Where a figure didn't survive a re-check, it isn't here.
| Project | Result | Conditions |
|---|---|---|
| tessl | 3.06–4.37× faster bf16 inference than PyTorch MPS; fits 4K–6K ragged workloads in ≤1,911 MiB where PyTorch OOMs at 33.6 GiB | Paired against sentence-transformers 6.1.0 / PyTorch 2.14.1 MPS across 6 EmbedGemma 2 workloads on M5 Pro (host load 5.1–16.7, 2026-10-21); pure-Rust SentencePiece BPE runs 2.2–3.7× faster (3.79M–5.12M tok/s, 4,671/4,671 exact ID match) than HF tokenizers 0.23.2. |
| ojas | Activation checkpointing cuts peak 18.08 → 3.96 GB (−78%); LoRA dW-skip cuts peak 11.81 → 6.63 GB (−44%) and step time −13.2% (6,314 → 5,480 ms) |
MetalBackend on M5 Pro, T = 2,048 (0.8B-Base checkpointing, n = 1, 28 GiB budget; 2B-Base dW-skip over 5 interleaved reps across 150 adapter sites, 2026-10-22). Metal Flash cross-attention also ran 34.5× faster (6.06 vs 209.28 ms) against ojas's own scalar kernel on SpikeTrack (2026-11-02). |
| ScholarLM | 97.7% (85/87) precision on public LitQA2 without a supplied paper corpus |
End-to-end literature pipeline on public LitQA2 queries with zero pre-supplied papers, improving over the closed-book baseline (2026-10-25). |
| DevMap | 1.6–21× faster cold index and 2.2–107× faster unchanged-tree refresh than four other code-graph tools; schema v29 cut index size 593.4 → 272.6 MB (−54%) |
Indexing speed measured on v0.2.2 (M5 Pro, fixed commits, lost single-file re-index to CodeGraph, 2026-09-14); schema v29 storage measured on the cold GitPulse corpus (2026-11-02). |
| GitPulse | 30k-commit graph refresh 25.19 → 1.11 ms; 24-watch interrupt wakeups 117.3 → 2.7/s; hook gate 32.16 → 11.16 ms (−65%); 100k-task search p95 97.59 → 44.60 ms | Revision-token graph refresh (4.79 → 1.15 ms at 5k across 21/21 rounds, 2026-11-09); kernel kqueue/wake-pipe loops on macOS (2026-10-13); local-socket hook gate (7 interleaved rounds, 2026-10-13); interleaved 100k-task search on M5 Pro (2026-10-07). Also cut verification sweep duty ~80% → ~27%. |
| Gusset | Go→Rust serial call 90.3 → 10.6 µs (8.5×) | Spin-then-park, certified before/after on one linux-amd64 VM (2026-10-08). A later shared-memory ring (7.46 → 3.73 µs) ran in a different container without a provenance header, so the two are not chained. |
Each project is a module the next one is built on. Every arrow below is a real dependency in the source: a crate, a go.mod require, a vendored engine or a spawned sidecar. Arrows point from a project to what it builds on.
flowchart TB
subgraph K["Kernels"]
tessl["tessl<br/>Metal 4 GEMM, Q4/Q8 + Qwen3.5 engine"]
sparsl["sparsl<br/>sparse + scan kernels"]
end
subgraph E["Engines & boundaries"]
ojas["ojas<br/>deep learning & training engine"]
gusset["Gusset<br/>Rust-in-Go contract"]
end
subgraph A["Code intelligence & agents"]
devmap["DevCouncil · DevMap<br/>code graph + gate"]
manvi["MANVI<br/>agent harness"]
jarvis["Jarvis<br/>model-free replay"]
end
subgraph P["Products"]
gitpulse["GitPulse"]
scholarlm["ScholarLM"]
devtype["DevType"]
vani["Vani"]
end
subgraph R["Research"]
binn["BINN"]
lappi["Lappi"]
gemma["gemma-metal"]
spiketrack["SpikeTrack"]
end
ojas -->|kernels| tessl
ojas -->|cgo| gusset
lappi -->|Mac backend| tessl
lappi -->|Mac trainer| ojas
binn -->|crates.io| sparsl
binn -.->|optional| tessl
gemma -->|GEMMs| tessl
spiketrack -->|Metal 4| tessl
spiketrack -->|runtime| ojas
devmap -->|engine + oracle| gusset
manvi --> devmap
manvi --> gusset
jarvis -->|replay| manvi
gitpulse -->|vendored| devmap
gitpulse -->|sidecar| manvi
gitpulse --> gusset
scholarlm -.->|dev tooling| devmap
scholarlm -.->|workspaces| gitpulse
devtype -.->|dev tooling| devmap
devtype -.->|coverage| gitpulse
vani -->|policy pkg| manvi
vani -.->|EmbedGemma 2| tessl
vani -.->|triage + labels| lappi
vani -.->|dev tooling| devmap
sparsl was lifted out of BINN's numeric core and published on its own. GitPulse and DevCouncil power development workflows across ScholarLM, DevType, Vani, and the rest of the workspace; GitPulse also vendors MarkDev's renderer crates, and DevPrism embeds MANVI as its tool gate. Explore the interactive stack on bharath.vbcr.dev.
| Project | Builds on · Used by | |
|---|---|---|
| Kernels | ||
| tessl · tessl.vbcr.dev Makes matrix math and LLM inference fast on Apple GPUs (v0.2.0 on crates.io). A Metal 4 GEMM and tensor runtime in Rust using MPP TensorOps matmul2d, cooperative register accumulators, fused epilogues, MLX-affine Q8/Q4 dequantization, and Metal 4 e-means kernels. Runs EmbedGemma 2 3.06–4.37× faster in bf16 than sentence-transformers on PyTorch MPS (fitting 4K–6K ragged inputs in ≤1,911 MiB where PyTorch OOMs at 33.6 GiB) and tokenizes at 3.79M–5.12M tok/s (2.2–3.7× HF tokenizers); retired its standalone training step to ojas (C ABI 14) while supplying ojas's Metal backward kernels. |
Used by ojas, Lappi, gemma-metal, BINN, SpikeTrack | |
| sparsl · Fast, reproducible kernels for spiking-network simulation: resident CSR SpMV, fused LIF membrane updates, a chunked prefix scan, and MTLSharedEvent zero-copy handoff with tessl (v0.3.0). A Device exists only for a backend that can actually execute, so results reproduce bit for bit and never misreport where they ran. |
Lifted out of BINN · used by BINN | |
| Engines & boundaries | ||
| ojas · ojas.vbcr.dev Trains and runs neural networks on Metal, CPU, and WebGPU inside Go services with no Python runtime. Owns the Metal/CPU reverse-mode training step: LoRA dW-skip cut Qwen3.5-2B device peak 11.81 → 6.63 GB (−44%) and step time 13.2%, activation checkpointing cut 0.8B peak 18.08 → 3.96 GB (−78%), Flash cross-attention ran 34.5× faster than its own scalar kernel on SpikeTrack, and 118 C-ABI checks pass alongside in-process Go tests. |
Builds on tessl, Gusset · used by Lappi, SpikeTrack | |
| Gusset · gusset.vbcr.dev Lets a Go service call a Rust engine safely and cheaply (v0.0.3, MIT / Apache-2.0): panic firewall, bounded uncollected completions (1,024 default), bounded 30 s Close with worker process-group teardown, deadlines enforced inside Rust, poisoned handles, per-field ABI checks, and ARM64 wait-ring formal concurrency scaffolding. |
Used by DevCouncil, MANVI, ojas, GitPulse | |
| Code intelligence & agents | ||
| DevCouncil · DevMap · devcouncil.vbcr.dev Gives AI coding agents a fast, accurate map and verification gate (v0.2.3). DevMap builds stacked zero-copy SQLite code graphs (schema v29 shrank cold GitPulse index size 593.4 → 272.6 MB), serves 11 MCP tools (devmap ask, blast radius with owners, commit suspects), and links Lean 4 TCB / anti-reward-hacking checks; write-policy matching runs in Go (Gusset handles gusset-check and the dc-glob oracle), with 111/117 Swift and 74/79 Python resolution fixtures passing. |
Builds on Gusset · used by MANVI, GitPulse, ScholarLM, DevType, Vani (dev tooling) | |
| MANVI · manvi.vbcr.dev Runs AI coding agents under explicit policy in pure Go (1.27.4) and Rust across a process boundary: 51 native tools (25 dynamically activated core + 26 extended), SQLite lease exclusion, Antigravity MCP support, and a six-step policy ladder whose outcomes stay distinct (1,031 cross-language parity cases, 86.9% Go statement coverage across 21 zero-tolerance linters). |
Builds on DevCouncil, Gusset · used by GitPulse, Jarvis, DevPrism | |
| Jarvis · jarvis.vbcr.dev Desktop capabilities discovered once with Gemini, frozen into typed artifacts, then replayed through MANVI with no model decisions: 35/40 macOS replays and 40/40 saved-state checks, not yet a clean stability pass. Human approval gates every account change. |
Builds on MANVI | |
| Products | ||
| GitPulse · gitpulse.vbcr.dev Single-process Rust / Tauri 2 / Svelte 5 control plane (zero telemetry) for Git review, worktree collision detection, task orchestration, and supervised AI agent sessions: split-pane canvas terminal multiplexer, native wry browser tabs with element picker and port radar, 22 MCP tools (11 DevMap code-graph + 11 worktree/task tools), per-user class-based git spawn budgeting, 6-phase git option-injection hardening, 322 contract-checked IPC handlers (2,080 struct fields), and 95.07% Vitest frontend coverage across 14,072 tests. |
Builds on DevCouncil, MANVI, Gusset · drives workspaces across ScholarLM & DevType | |
| ScholarLM · showcase & architecture AI research platform searching 20 scholarly sources across ~250 Go routes and writing fully-cited, grounded manuscripts (plan → write → verify → review, 97.7% [85/87] precision on public LitQA2 with no supplied corpus): React 19 PWA, Rust edge (with MCP OAuth 2.1 / PKCE S256), Go orchestration core, Python ML worker, and a 25-tool MCP server whose agent layer is open as WisDev. |
DevMap & GitPulse power its coding-agent workflows | |
| DevType · devtype.vbcr.dev Native macOS text expander and on-device writing assistant in Swift/AppKit (v1.2.0 + companion devtype-cli, 5,085 automated tests executed). Typed triggers expand in ordinary text fields; proofread, rewrite, translate and code actions run on Apple Foundation Models offline; imports TextExpander/Espanso and protects secrets behind Touch ID. |
DevCouncil verifies its changes; its coverage export feeds GitPulse (development tooling) | |
| Vani · architecture & apply suite Listings-first career engine and local-first apply suite in Go ( vanid), Swift (macOS Vani Apply), and TypeScript (Chrome/Safari WebExtensions). Ingests roles across 30 ATS platforms, drops stale/duplicate listings at zero token cost, triages open roles across four slots (role fit, sponsorship, work mode, seniority), exports BLAKE2b MinHash held-out fingerprints so Lappi's contamination gate refuses train/eval overlap, and generates grounded rkb-verified resumes behind a fail-closed claim eligibility gate. |
Builds on MANVI, tessl (EmbedGemma 2), Lappi, DevCouncil tooling | |
| Research | ||
| 🧠 | BINN · binn.vbcr.dev From-scratch Rust instrument that falsified backprop-free local learning under pre-registered kill-gates, then earned a positive on Spiking Heidelberg Digits (0.8320, 12/12 seeds ≥ 0.80) proving temporal spike order drives the causal-attention read-out; bit-exact NEON attention (1.85× fewer cycles) and Accelerate BLAS (2.23× fewer cycles) accelerate 2-epoch SHD cells over sparsl 0.3 resident ticks. |
Builds on sparsl, tessl |
| SpikeTrack · architecture & parity Pure-Rust port of Meta's SAM 3.1 multiplex video tracker on Apple Silicon (zero Python at runtime, Pillow-exact JPEG decoder, 14 linked system frameworks): mixes an fp16 Core ML default trunk (0.569–0.600× fp32 trunk frame time on VISEM) with merge encodes, segmentation head, and Metal 4 Flash cross-attention on ojas / tessl (0.5695 held-out HOTA vs 0.5700 fp32 across 1,500 frames). |
Builds on ojas, tessl | |
| Lappi · lappi.vbcr.dev Calibrated typed-decision model (v0.1 preview 0.857 choice top-1, not promoted after OOD abstention 149/180 missed its 90% bar): schema in, typed slots out (choice, score, span, abstain) where not_run never counts as a pass. Now serves calibrated job-modeling heads (role_fit, work_mode, seniority, sponsorship) on ojas / tessl behind a MinHash contamination gate and −47% scan peak containment (20.6–22.6 → 10.9–12.0 GiB). |
Builds on tessl, ojas | |
| nanolab · attention.vbcr.dev Instrumented small-LM training lab (378/378 checks passing): attention, Mamba-2, Gated DeltaNet and minGRU behind CLI flags, chunk-parallel scan kernels, and seed-paired intervals — withdrawing unpaired GDN/Mamba-2 speedups directly in the committed TMLR/preprint manuscript. |
Research companion to the kernels | |
| gemma-metal Gemma inference runtime for Apple silicon with split sliding/global KV ring caches. Takes its general and INT4 GEMM kernels from tessl, and remains below its own decode-speed gate (6.83 vs 15 tok/s on 31B Q4; 23.9 vs 48 on E4B). |
Builds on tessl | |
These are directions, each grounded in what the repositories themselves record as open:
- Train and serve Lappi on my own stack. Lappi's Mac trainer already runs on
ojasand its backend serves ontessl's Qwen3.5 kernels, while the main2Bcampaign runs on PyTorch. Next: OOD abstention, serving the code-defect family, and promoting job-modeling runs past pre-registered validation floors. - A device-aware engine for Go services.
ojasreaches Go through Gusset, and its resource plan reads cores, caches, unified memory, and cgroup limits — but there is no automatic device router yet. Routing work between CPU and Metal from that plan is next. - Spiking read-outs on the current kernels. BINN's
tesslinterop is pinned to0.1.4, whiletesslhas moved to0.2.0(C ABI 14). Bringing the attention read-out that earned the SHD result onto the current kernels comes next. - One shared code-graph and policy substrate. DevMap and GitPulse already power engineering workflows across ScholarLM, DevType, and Vani, and MANVI under GitPulse, Jarvis, Vani, and DevPrism, so schema (
v29/v99) and formal-verification upgrades propagate across every repository at once.
| Paper | Journal | Year |
|---|---|---|
| Investigation on the heating effects of intra-tumoral injectable magnetic hydrogels (IT-MG) for cancer hyperthermia | Biomedical Physics & Engineering Express | 2025 |
| The Therapeutic Scope of Orofacial Mesenchymal Stem Cells | Bioengineering | 2025 |
Biomedical computing: GenoThermal_Targeting, a patient-specific magnetic-nanoparticle therapy pipeline from genomic discovery through physics simulation. Research briefs are at research.vbcr.dev.
Side projects: finished or maintained, but outside the main stack
- Graft: sealed Mac-to-Android continuity workspace (clipboard images/text, zipped folders, notifications, remote shell and audio) over mutual TLS 1.3 and Tailscale/USB, checked by a symbolic Dolev-Yao verifier, Kani proofs, and Lean 4 protocol specs.
- Chronicle: local-first second brain across Mac and Android (pinned LAN TLS, E2EE locked-vault parking, local Ollama RAG).
- MarkDev: native macOS Markdown editor (
v0.0.3) on a Swift + 4-crate Rust core; its renderer crates are vendored into GitPulse. - DevPrism: local-first LaTeX and research workspace, forked from claude-prism, with MANVI as its tool gate.
- M5Blade: Apple-silicon fan controller that writes to the SMC behind a race-free control gate.
- Strait: macOS bulk transfer (
F_NOCACHEI/O, pure-Swift BLAKE3 hash-on-write, and resumable staging). - Curio, ChronosFlow, Meridian, TrailBlazer: on-device AI and native mobile apps.
- SalEdge: multi-firm ERP for battery retailers, with GST e-invoicing and a local AI layer.
- AcademiaTrack, Void, Whimsical-Love: web apps and experiences.
Everything has a page at apps.vbcr.dev.


