01
SPR Biosensor Comparative Approach
A multilayer transfer-matrix simulation comparing sensor stacks, cross-validated against an established package with the disagreement from the cited reference documented explicitly.
Companion pageI build applied software and AI tools around real problems I can inspect.
I came to software through electrical and electronics engineering. The problems I enjoy most are the messy, real-world ones: making sense of road footage and weather signals, or protecting files that need to stay private. To me, the work matters only when it helps someone make a better decision, avoid a mistake, or keep something important safe.
Building now
Road Risk Review is a private product beta for turning recorded road footage into event timelines and investigation packs. Its public proof pairs a reviewed positive with a retained negative control using BADAS-Open on a pinned V-JEPA 2 backbone, measured NVIDIA L4 compute, and a causal virtual-clock projection.
Active product
Road Risk Review
A private beta for turning recorded road footage into causal model replays, evaluation gates, and reviewable investigation artifacts.
Reviewed causal replayBADAS-Open 1.0.0V-JEPA 2Measured NVIDIA L4 computeVirtual-clock projection
Observed outcome
Dataset positive: policy confirms
The exploratory policy confirms at 4.133 s. This is a recorded causal replay, not a live warning.
Boundary: measured-latency virtual-clock projection. Scores are uncalibrated. Human review required. Not ADAS.
Inspect causal replay manifest- 0 / 129
- Raw 8 Hz samples meeting the 125 ms compute deadline
- 4.00–4.02 Hz
- Measured-latency virtual-clock projection output rate
- 1 false positive
- Frozen exploratory policy confirmation on the negative control
The replay aligns BADAS-Open/V-JEPA 2 collision-class scores with a score-blind latest-window scheduler driven by measured NVIDIA L4 compute latency. It displays only samples available by the current source time and never interpolates from a future model result.
The paired evidence keeps both outcomes visible: the policy confirms the labeled positive at 4.133 s and also confirms the labeled negative at 4.867 s. Publishing that same-policy false positive makes the current failure mode inspectable.
Second product
Skywatch
A live weather prototype built around inspectable data.
Skywatch combines a seven-day forecast with server-side location privacy and an optional public-webcam cross-check. The companion page makes provenance, handling, and the visual check inspectable.
Inspectable path
v0.1 prototype- 01
Location
Resolved server-side; request IP discarded after lookup
- 02
Forecast
Open-Meteo data with its source and issue time visible
- 03
Visual check
Public webcam frames compared with brightness and edge signals
Open source
Sindook
Post-quantum file encryption built on X-Wing.
Sindook seals files and secrets against harvest now, decrypt later attacks. Key establishment combines X25519 with ML-KEM-768 through the X-Wing hybrid KEM, implemented from the IETF specification in Go with zero custom primitives. Sealed files carry key slots, so access can be rotated and old files upgraded in place without re-encrypting the data they protect. The CLI covers the whole lifecycle: sealing, verifying backups without writing plaintext, inspecting files without credentials, and rotating access across a directory in one command. Sindook is the Bengali word for a strongbox.
Verification path
v0.4.0- 01
Test vectors and interop
Verified byte for byte against the IETF draft vectors, and cross-tested against Cloudflare's and Filippo Valsorda's implementations on every CI run
- 02
Primitives
ML-KEM-768, X25519, SHA-3, ChaCha20-Poly1305, HKDF, and Argon2id, all from the Go standard library and x/crypto, none custom
- 03
Key rotation
rewrap swaps recipients by rewriting only the header, payload bytes untouched; a deep mode re-encrypts when someone must actually lose access
- 04
Tamper suite
Truncation, reordering, bit flips, wrong keys, slot stripping, and parameter downgrades each fail authentication
- 05
Adversarial testing
Five fuzz targets across header parsing, payload streaming, and KEM decapsulation, around 14 million executions with zero findings; a paired timing test guards the decapsulation rejection path on every CI run
Technical directions
- 01
Applied AI for real-world review and operations workflows
- 02
Computer vision on recorded road and environmental footage
- 03
Backend pipelines, evaluation gates, and evidence exports
- 04
Post-quantum cryptography and data protection tooling
Archived projects
Earlier scientific computing work, kept because it already carried the habit the newer projects are built on: cross-validate against an established reference and document the disagreement instead of hiding it.
Open to
Remote or relocation-friendly opportunities in applied AI, backend, computer vision, and product engineering, especially tools for review, analysis, or operations workflows.
Start a conversation