Delhi, India
Software Engineer, Dealermatix
Hello, I'm
Kartik
Verma.
Systems built to hold.
I build products that stay correct under pressure: money spent safely, one winner in every race, and stolen tokens that are worthless.
push it.
it holds.
Designed for the failure case,
not just the demo.
Tests behind
Sweep
Job boards in
one sweep
Fewer production
API failures
Selected work
Sweep,
end to end.
A real run on my own résumé, on the live site. Drag the strip to follow it from reading the résumé to the shortlist.






free job boards, plus LinkedIn, Indeed and Naukri
jobs actually applyable on one sweep, after reachability filters
tests across the engine, web app and workers
to read a résumé on a scale-to-zero GPU, against 187s on CPU
Spending other people's money safely.
Paid searches run on each visitor's own Apify key. Every run carries a cost cap the provider itself enforces, the cap is reserved before the run starts, and if the client can't set the cap the run is refused.
# scraper.py: fail closed, not open
ceiling = max_charge_usd(site_key, effective["max_results"])
if ceiling is not None:
if not charge_ceiling_supported(actor):
# starting anyway is the run this guard exists to bound
raise RuntimeError(f"{site_key}: cannot set a charge cap")
start_kwargs["max_total_charge_usd"] = ceilingThe 90
seconds.
NearMed: a customer sends one prescription photo, nearby pharmacies each send a price, and the customer sees one price at a time.





No timers. One winner.
Every change to a prescription's queue runs under a row lock. Deadlines are worked out whenever someone reads the queue, and two partial unique indexes allow one price on screen and one accepted price. A retried tap gets the same order back.
# offers.py: expiry decided on read
def current_offer(s, rx, now):
q = first(in_line(s, rx.id, "showing"))
if q and now > q.deadline + GRACE:
q.status = "timed_out"; q = None
if q is None and is_open(rx, now):
q = first(in_line(s, rx.id, "queued"))
q.status, q.deadline = "showing", now + OFFER
Renting between
strangers.
RentKaro is a peer-to-peer rental app. The money and the sessions are built so neither side has to trust the other.
- A stolen token, made worthlessReplaying a refresh token that already rotated revokes the session and invalidates every access token.
- Money that can't overdrawConditional wallet updates inside one transaction; the deposit stays held until an OTP-confirmed return.
- An account takeover, found and fixedPassword reset now needs a purpose-bound, encrypted email code.






Before.
After.
Restory looks at each old photo, picks GFPGAN, CodeFormer or a two-pass chain, and tells you why. This is an 1890 portrait, restored on a laptop CPU.
"small/low-quality faces → CodeFormer (robust)"


The work.
The proof.
Numbers from production, from test suites and from real runs. Each one traces back to a project above.
measured
outcomes
- 4,577
Tests behind Sweep
Engine, web app, workers and benchmarks
- 37/727
Jobs actually applyable on one sweep
Sweep: reachability and repost-farm filters
- 1
Price on screen, and one winner
NearMed: row locks and partial unique indexes
- 40%
Fewer production API failures
Dealermatix integration layer
- 85%+
Apex test coverage on new features
Dealermatix
- 9s
To restore an old portrait on a laptop CPU
Restory, with automatic model choice
I'm drawn to the engineering nobody sees.
Full-stack engineer in Delhi with about two years building production mobile and web products. What grips me is concurrency, auth and the failure modes that keep real products from breaking under real load.
Education
Where I first worked with generative models, the thread that later became Restory.
How I work
- 01Find the failure
Start from how it breaks: the race, the replay, the runaway bill.
- 02Model the state
Make every state explicit so nothing lives in a gap.
- 03Guard the boundary
Validate at the edge and fail closed.
- 04Test the race
Replay the timeline with fixed clocks before a user ever sees it.
- 05Ship measured
Reliability you can point a number at.
Stack
The race that never happens and the stolen token that's worthless are the wins I care about most.
KartikJourney
Promoted from Trainee. Salesforce and React Native integration work on a live dealership CRM: 5+ REST integrations, OAuth 2.0 across 20+ endpoints, Apex and LWC, 85%+ test coverage on new features.
Designed, built and launched a job search engine: a self-hosted model reads the résumé, 133 job boards are searched, and paid searches stay inside provider-enforced caps.
A pharmacy price queue with exactly one winner, a rental marketplace with a deposit-holding wallet, and a photo restorer that explains its choices.
Started on an existing, load-bearing Salesforce codebase, inside strict API limits and someone else's conventions.
Where I first worked with generative models.
Milestones
- Sweep launched as a public beta
- Promoted from Trainee to Software Engineer
- NearMed's price queue: one price on screen, one winner
- Found and fixed an account takeover in RentKaro
Let's build
something
that holds.
Email is the fastest way to reach me.
Running Sweep's public beta on the side.