X:0000 Y:0000
GK/SPEC-2026
Inspected & Approved
DOC GK-001 REV 2026.07 CLASS FAILURE-AWARE SYSTEMS UPTIME 00:00:00 EDITION

Systems that keep running when the network doesn't.

I'm Gautam, Computer Engineering @ uOttawa. I build software around the assumption that networks drop, APIs die, and hardware misbehaves. Don't take my word for it: flip the UPLINK switch in the corner and watch this page handle its own outage. Try it

400 PX
PERSONNEL REF. 01
↖ DRAG ME
LOCATIONOttawa, ON · Canada
STATUSOpen to co-op
FOCUSEdge AI · Platform · Infra
PRIMARYC++20 · TypeScript · Python
DEPENDENCIES0
SEC. 00

Live simulation · AEGIS

AEGIS-SIM · CAM-01 · ZONE A ● REC SSD-MOBILENET · 300×300 · ON-DEVICE
SAFETY STATE
SAFE
RELAY / GPIO
IDLE
EVENTS UPLINKED
0
OFFLINE BUFFER
0

This is the real AEGIS pipeline logic: detection → zone rules → state machine → alarm → async uplink, reimplemented in ~150 lines of browser JS. Workers without hard hats trigger a violation when they enter the hazard zone; the relay fires before the voice, exactly like the SCHED_FIFO ordering on the device. Now flip the UPLINK switch above and watch events buffer offline, then flush when the link returns. Nothing in the safety loop notices. Full design decisions in the engineering log. And the switch is not just a toy: turn your wifi off and reload this page. It still loads, and the switch flips itself.

01
SEC. 01

Selected work

SHOWCASE A/F · EDGE / RTOS

AEGIS EDGE · SURVIVES OUTAGE

The safety loop never touches the network

C++20 on QNX 8.0, four processes over MsgSend/MsgReceive, detection to relay entirely on-device. The dashed uplink box is the point: it's the only optional part.

READ THE ENGINEERING LOG →
SHOWCASE B/F · AGENT SECURITY

mcp-pin SIGNED PUBLIC LOG

The tool you approved is not the tool you're running

You approve an MCP server's tools once and that decision is never revisited, even though a tool description is an instruction with the reach of a system prompt. mcp-pin fingerprints what you approved, blocks the session when it moves, and keeps a signed public history of every version across 248 servers. Its own recrawl corrected 17 of its earlier numbers.

READ THE ENGINEERING LOG →
SHOWCASE C/F · NUMERICS / PUBLISHED

assay-gpu PEER-SHAREABLE · DOI

The detector that failed its own test

Silicon sometimes returns a wrong answer silently. The 1984 checksum that catches it has a threshold that is undefined at bfloat16, so I measured the distribution instead and calibrated one. It catches 76% of exponent-bit flips against a 90% bar I registered first, and none of 600 low-mantissa flips. Four predictions failed, four findings retracted, all published.

READ THE ENGINEERING LOG →
SHOWCASE D/F · SUPPLY CHAIN / PUBLISHED

LADING ARCHIVED · DOI

It failed three times, and the third found the other two

Regulators now ask which known flaws actually reach your product. I built the tool that answers it from the shipped binary. 99.6% of findings stopped before a binary was ever opened, every clearance it did make was unfalsifiable, and the fix for that was aimed one layer too shallow. Both kill tests resolved against it, so the project stopped. The paper, the dataset and two upstream bug reports are the output.

READ THE ENGINEERING LOG →
SHOWCASE E/F · AGENT EVALUATION

Vernier CONTROL FIRST

Temperature 0 is not deterministic

A deterministic diff for agent behaviour: record a run, re-drive it after a change, get a rule-based verdict with no LLM judge in the path. Before measuring anything I ran the control, and 6.7% of runs diverged with every knob pinned. My first version of the study was wrong; the retraction is still in the repo.

READ THE ENGINEERING LOG →
SHOWCASE F/F · ROBOTICS / SAFETY

Cairn EDGE · UNDER 1 ms

Catching a robot that is confidently lost

GPS dies underground and SLAM fails silently: the pose drifts while the covariance stays tight. Cairn is the error bar, not the dot. On injected drift a naive monitor never alarmed and error hit 20.5 m; cross-source consistency flagged it in 12 s at 0.54 m. Its own worst case, common-mode faults at 73% missed detection, is published in the README.

READ THE ENGINEERING LOG →
A-01

AeroGuard

Confidence-scored SOC incident triage, dual-store: MongoDB for live reads, Snowflake for analytics, so analytics never contend with real-time state. 7-test E2E suite, shipped in 24 hours.

PYTHONFASTAPISNOWFLAKEMONGODB
A-02

Ephemera

Open a PR, get a full stack; close it and it's gone. A webhook writes desired state only, and a crash-safe reconciler advances one step per tick against idempotent providers.

TYPESCRIPTHONOBULLMQPOSTGRES
A-03

HyperShift

Five agents turn plain English into deployed infrastructure. Approval gates, rollback, immutable audit log. Nothing reaches prod without a human.

5 AGENTSBULLMQZODTURBOREPO
A-04

MetaShift

Multi-tenant SaaS end to end: Postgres RLS, JWT and API keys, live conflict resolution over WebSockets, metering, billing, audit.

POSTGRES RLSWEBSOCKETSSTRIPEMETERING
A-05

Groundhog

Ranks the chores you keep paying an agent to redo, from local session history that never leaves the machine. It refuses to render a partial scan as a clean null, and every missing file is named in a NOT COUNTED ledger.

PYTHONLOCAL ONLYSTDLIB ONLYSELF-CHECK
A-06

Chronos

Digital-twin crisis management over 14 live APIs on NATS. Led a team of 4 at uOttaHack; outages degrade, never kill.

NATS14 LIVE APISDIGITAL TWINTEAM OF 4
A-07

Bedrock

Every scraper heals itself with AI now; nobody checks the repair was correct. Bedrock verifies each heal against golden fixtures before it ships, and every row carries a receipt naming the heals it refused.

NEXT.JSPOSTGRESMCPBRIGHT DATA
A-08

keyring

Access governance for people and agents. Finds every grant still held, holds it at a human approval gate, and writes a verifiable audit trail of what was granted and by whom.

TYPESCRIPTMCPAUDIT LOGAPPROVAL GATES
COMMENDATIONS & PUBLICATIONSJUDGED OR PUBLISHED · NOT SELF-REPORTED
PLIMSOLL Pre-Registration Specification v1
Technical note · DOI 10.5281/zenodo.22107451
PLIMSOLL · released CC0 so anyone can implement it
Binary Grounded VEX Clearance Fails Three Times
Second paper · DOI 10.5281/zenodo.22099794
LADING · separate dataset DOI, 55 artifacts, hand labelled ground truth
Calibrating the Checksum
First paper · DOI 10.5281/zenodo.22054179
assay-gpu · CC BY 4.0, data and pre-registrations released
Best Use of Snowflake API
MLH · HackConcordia
Credits Pool Winner
The Zerops Challenge · WeMakeDevs
Ephemera · $1,000 in credits, one of five
0
ENGINEERING LOGS
0
PUBLIC REPOS · CACHED
0
OPEN-SOURCE RELEASES
0
NETWORK DEPS IN SAFETY LOOP
02
SEC. 02

Operating principles

P-01

Design for the outage first

Offline buffering, graceful degradation, and chaos harnesses go in the initial plan, not the retro.

PROVEN · GK-002 §05: the chaos harness that caught the lossy flush
P-02

Cloud makes it smarter, never dependent

Enrichment and dashboards live upstream. Correctness lives on-device, where the network can't take it away.

PROVEN · GK-002 §02: the voice baked in at build time
P-03

Autonomy needs a leash

Approval gates, audit trails, rollback. Every autonomous system I ship can be stopped, questioned, and reversed.

PROVEN · HyperShift: nothing deploys past the approval gate
P-04

Prove it without the hardware

Mock generators, load testers, fault injection, so the stack ships while the Pi is still in the mail.

PROVEN · GK-002 §05: the upper stack shipped before the Pi arrived
P-05

Publish the negative result

A study that only reports its wins is not evidence. I pre-register the threshold, then report the number I actually got.

PROVEN · GK-005: needed 20%, got 9.8%, published in full
P-06

Caveat louder than you claim

A measurement without its conditions is a rumour. If the harness distorts the number, that warning goes above the number, not in a footnote.

PROVEN · GK-003: WSL2-relative timings, flagged before the table
P-07

Retract it when it's wrong

Clearing your own pre-registered bar by five times is an alarm, not a triumph. The broken write-up stays published, marked do-not-cite, because deleting it makes the repo look better and the record worse.

PROVEN · GK-006: the 100% that turned out to be a dead hostname
P-08

Write the kill test before the code

Name the number that ends the project while you still have nothing invested in the answer. Two of mine are still sitting at UNEVALUATED in public.

PROVEN · GK-004: KT-1 and KT-2 committed before the first commit
P-09

The refusal is the product

A gate with an override is decorative. If a sealed rule can be amended by a flag, a config or a paid tier, the artifact is worth nothing to the person it was meant to convince.

PROVEN · GK-010: a test that fails if a code path to mutate a sealed field ever appears
P-10

Stop when your own rule says stop

A stop condition you argue your way past was never a stop condition. Three not-evaluable readings, and the pre-registered action was to archive it.

PROVEN · GK-009: both kill tests resolved against it, so the project ended
P-11

Name the competitor who did it better

Finding prior art that beats you on the job it does is uncomfortable and burying it is cheap. A security tool that hides the alternative has already told you what kind of tool it is.

PROVEN · GK-011 §06: the README recommends mcp-warden over my own project
P-12

Store the conditions with the measurement

A number whose environment was not recorded cannot be re-derived, only re-argued. I have written that caveat on other people's results and then failed to apply it to my own crawl.

PROVEN · GK-011 §08: the recrawl that disproved my own explanation
03
SEC. 03

Field record

FILE 03-A JUL – AUG 2026 ● ACTIVE DUTY

Data Processor & Software Developer

Canadian Musicians Cooperative

Built Cadence, a production inventory and rental system, solo, from requirements gathering with non-technical staff to deployed software. Audited and repaired data-integrity failures and automated recurring workflows.

NEXT.JSPRISMAPOSTGRESQLSTRIPERBACEN / FR
SHIPPED SOLO
FILE 03-B OCT 2025 – JAN 2026 TOUR COMPLETE

Financial Sales Representative

Kognitive Sales Solutions × Canadian Tire

Explained regulated financial products to high-volume customers under strict compliance: daily practice at making complicated things understandable to people who didn't ask for the details.

COMPLIANCEHIGH VOLUMEPLAIN ENGLISH
FIELD TESTED
FILE 03-C 2024 – PRESENT ● IN PROGRESS

B.A.Sc. Computer Engineering (Co-op)

University of Ottawa

Data structures & algorithms, databases, software engineering, computer architecture, discrete math, probability & statistics.

DSADATABASESCOMP ARCHDISCRETE MATHPROB & STATS
TRAINING
04
SEC. 04

Contact

The uplink is live.

Open to co-op placements and internships in systems, infrastructure, and platform engineering. If you're building something that has to survive failure, I want to hear about it. (Email works even during the outage; it's store-and-forward. That's the point.)

SAFETY LOOP · NOMINAL
GK-001 console. Type help.
gk@edge:~$
ADDRESS COPIED · developwith.gt@gmail.com