Harsh

Building AI reliability tooling · 7 merged PRs @ Agenta/TraceRoot/LiteLLM/Mastra
Node · Python · FastAPI · ML
MCA @ NIT Jamshedpur · open to remote roles

[Harsh @unfitcoder]

Day 241

Read about the “four burners” idea: family, friends, health and work. You can’t run all four at full flame, so you choose.

Right now work is the one I’m turning up. Tonight is system design.

Rejection stung, but it gave me a clear to-do list.
#buildinpublic

[Harsh @unfitcoder]

Day 240

Missed posting yesterday. An old client came back for website maintenance work, and I got rejected at the final round of a company interview.

Not going to spin the rejection. I’m going through where I lost marks and fixing those gaps.

#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Day 239

Pattern recognition assignment and Linux lab work today, for 10 marks. Not glamorous, and the gateway build hasn’t started.

Degree work and the build are fighting for the same hours. Next: one focused hour a day on the gateway.

#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Then I retrained 5 models on it. Ensemble AUC 0.516 vs 0.518 for a simple low-volatility rule. It beat the best baseline in 2 of 8 yearly folds.

So: no proven edge. I’d rather post that than a number I can’t defend.

#buildinpublic #MachineLearning

[Harsh @unfitcoder]

Day 238

Found a flaw in my stock model’s target. I labelled the top 30% of 10-day returns as winners. But volatile stocks hit extreme returns more often, so a model could look predictive just by spotting volatility.

[Harsh @unfitcoder]

The test: what share of winners sit in the most volatile third of stocks? (Neutral = 33.3%)

→ Original target: 40.5%
→ Volatility-adjusted target: 32.2%
The new target mostly removed the bias.

[Harsh @unfitcoder]

@shivxni_v Listen Everybody, here is the secret ☝🏼

[Harsh @unfitcoder]

One design rule I settled on: a wrong API key is never retried or failed over. A timeout or a rate limit can fix itself. A bad key won’t, so hiding it just makes it harder to find.

First curl result tomorrow.

#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Day 237

Scoped the first 3 hours of my LLM gateway: one endpoint, 3 providers, and a header that shows which one answered.

[Harsh @unfitcoder]

Day 236

Back on track. Solved a few DSA problems today.

Easing back in with a smaller target and no catch-up grind. Consistency matters more than making up for a rest day.

#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Day 235

Missed posting yesterday. Burnout caught up with me, so I took the day off: pizza, a trek, no screens.

Rest isn’t a skipped day. It’s what the next ten good days run on.

#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Not a big-output day, but clear boundaries now save a lot of cleanup later.

#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Day 234

Spent today getting my stock-signal project ready to submit for college.

I also caught myself about to mix it up with the new sentiment project, so I decided to keep the two in separate repos.

[Harsh @unfitcoder]

The part I like: I tested it with 16 threads, then removed the lock on purpose to confirm the tests fail without it. A test that can’t catch the bug isn’t a test.

Also added same-origin protection on the dashboard’s write requests. 212 offline backend tests passing.

[Harsh @unfitcoder]

Day 233

Missed posting yesterday, so here’s the catch-up. I spent the day hardening my stock-signal project.

Fixed a thread-safety bug in the price-data cache. Reads and writes from several threads could collide, so I added a shared lock and a 500-entry limit.

[Harsh @unfitcoder]

First goal: one endpoint that shows which provider answered.
#100DaysOfCode #buildinpublic

[Harsh @unfitcoder]

Day 232

No code written today. I spent it choosing and understanding the gateway build, because I’d rather know why I’m building each part than copy a tutorial.

Compared 20+ project ideas and picked the one that fills my real gaps: Redis, metrics and system design.

[Harsh @unfitcoder]

What learned:
→ A circuit breaker has 3 states: closed (healthy), open (stop sending traffic), half-open (test with a little traffic). Half-open is the “self-healing” part.
→ Every request will carry a tenant, a feature name and a request ID, so cost and failures can be traced.

[Harsh @unfitcoder]

@boardyai Ha, fair. Counting up is more fun. Day 231 now, gateway build starts today.

[Harsh @unfitcoder]

Day 230

Starting a 5-7 day build: a self-healing LLM gateway.

Problem: if your app depends on one AI provider, their outage is your outage. A rate limit or a slow spell breaks your feature.