Waqas Sharif

Built the evaluation harness the platform is maintained through: five named personas and a ten-scenario curriculum matrix producing 60 per-turn traces, 19 behavioural probes using paired before-and-after measurement and single-criterion isolation to attribute cause, a separated run-and-analyse pipeline, and a 46-file behavioural test suite of 592 automated tests catching regressions before users hit them.

Refuted my own diagnosis of why the system over-asserted. I assumed thin data caused it; measuring across five data-completeness levels showed the opposite, the thinnest persona was the most honest, declining outright rather than fabricating. The real trigger was internal signal divergence, the model inventing an affective bridge to reconcile conflicting academic and psychometric inputs, so the scarcity guard I had planned would have fixed nothing. Redirected the fix to the divergence path, then added a deterministic post-generation gate as backstop after establishing that negative directives carry a reliability ceiling.

Architected a 24-node LangGraph pipeline: fan-out across 9 concurrent nodes, fan-in through a central dispatcher, then two further stages, with 52 typed state fields and concurrent-write reducers. Cut first-response latency from 20 to 25 seconds down to 3.7 seconds. Per-node model routing lets any node's LLM be swapped through environment config alone, with LangSmith tracing for per-turn execution, node latency, and token usage.

Built a production FastAPI backend: 22 REST endpoints, LangGraph checkpoint persistence to PostgreSQL, SSE streaming with a parallel keepalive task that defeats the host's 30-second proxy timeout, JWT authentication with bcrypt, rate limiting, 11 Alembic migrations against live data, a deterministic refresh endpoint that regenerates recommendations with zero LLM calls, and a profile summary generator running six parallel LLM calls plus synthesis off the event loop. Automated a 15-day job-market pipeline mapping 14,595 job titles to canonical career fields via Gemini.

Engineered two model-independent LLM guardrails: a 9-criterion reflection gate that applies surgical corrections before any response streams, including a check that blocks unfounded emotional or clinical claims about a user, and deterministic context stripping, built after finding the model ignores "do not include" instructions but cannot discuss data it never receives. Verified across two model versions, alongside three callable tool interfaces and PII handling kept as concerns separate from the agents themselves.

Shipped a 29-route Flutter frontend: onboarding, six assessment flows, O/A Level grade conversion, live streaming chat, recommendation dashboard, and a profile hub. Riverpod state management and cold-start timeout handling. Web build deployed on Vercel (live), with a contract test on the frontend asserting the same merit-tier contract the backend enforces.

Designed the recommendation engine: a 265-item session drawn from item banks of 1,409, sized so a student who reassesses is not asked the same questions twice, across seven construct-grounded psychometric instruments including an original one derived from work-adjustment theory, with MBTI rejected on published reliability evidence. Scored by a multi-dimensional model blending interest-profile geometry, measured capability, and deterministic tier validation with no LLM call. Fully bilingual: every instrument and the entire session exist in English and Roman Urdu, and the conversation follows whichever the student uses, including mixed script within a sentence.

Ran the build through a two-tier development harness of my own design, now my default method for any system of size: an architecture conversation producing versioned specifications, an implementer working from them under a written rules file across five role-scoped instruction sets, and LangSmith traces pulled back through a fetch skill to close the loop. Recorded across 334 session records in four months with the model and reasoning tier attributed per session, 53 audit documents over six whole-system audit phases, and a status register tagging every claim live-verified or believed so that doubt stays pointed at the unverified set.

Coordinated two contributors through written setup guides, versioned interface contracts and 17 dated change notices, against a governance document versioned to v2.27 carrying 80+ locked decisions, with rehearsal runs before every destructive data operation and a documentation toolchain generating both the architecture view and the dissertation itself from code.

Identified and named three failure modes: Optimal Answer Trap (the model locks onto its first acceptable answer and resists valid redirection), Final Synthesis Bias (output collapses toward a synthesised endpoint regardless of instruction), and Conversational Gravity (accumulated context progressively narrows model behaviour), each reproduced through documented prompt sequences.

Ran controlled experiments across temperature, top-p, and top-k sampling; 1M-token context stress tests; adversarial red-teaming with hallucination injection; and context-ordering trials showing a ~30% retrieval error rate when context follows rather than precedes the instruction.

Shipped a 5-stage prompting protocol combining intent decomposition, strict output schemas, and adversarial self-critique, deployed as an interactive Streamlit application with a failure-mode comparison view and a prompt template playground (live).

Engineered a sequential-parallel-loop agent architecture: an intent synthesiser compacts chat history into a research directive, three live-search agents fan out concurrently, and a draft-critique-validate loop exits on an explicit escalation signal rather than a timeout.

Implemented per-call session isolation. A fresh session per pipeline invocation keeps raw search results, internal drafts, and critique logs out of the user's conversation context, a production concern absent from typical multi-agent demos.

Deployed a deterministic code-based keyword filter ahead of all tool access, making the security guardrail structurally bypass-resistant, with a deliberate worker/critic model split for quality-cost balance.

Cut job discovery from a by-hand baseline I estimated at three to six days to a measured median of 5.4 hours after publication, 90th percentile 10.5 hours over 114 postings, against targets of 24 and 72. The pipeline runs on GitHub Actions and Python (standard library plus requests), reading eleven employer ATS boards twice a day and one aggregator daily.

30 production runs over 13.6 days with zero retries and zero failed requests, a median of 36 requests per run against a budget of 500, and 34,557 postings fetched. Retry with backoff, a circuit breaker and three exit codes (done, failed, stopped and resumable) mean a missed run costs latency, never data.

Every posting is admitted or dropped by a named deterministic rule, with no model and no scoring, and every drop is logged. In one run of 1,283 postings, 36 were kept: 922 dropped on title against 79 terms in four role families, 258 on location, 56 on seniority and 11 on age. The location rule dropped exactly the 74 of 91 saved postings its decision record had predicted nine days before it was built. Duplicates are grouped: 824 postings became 527 display rows, and one role posted 134 times, once per city, became one.

Git branches serve as the database, in three layers: a permanent append-only raw layer, a filtered layer, and an Airtable display that is never authoritative. Privacy is enforced in code rather than by care: employer description text and aggregator data never reach the public repository, and every posting is saved in full to a private store, marked saved only after it is read back and matched by content hash. A daily contract check fingerprints each API's field shapes, telling a board that changed its API apart from one with no new jobs.

Ran the build through a four-role AI-native development method of my own design. I bring the blueprint, the architecture and the engineering trade-offs, and set the standards every role works to; an architecture chat refines the design with me into decisions; Claude Code implements to briefs I approve; a fresh read-only reviewer audits cold; and every brief and audit comes back to me to decide. Six independent audits, 55 decision records, 23 session logs and 122 commits on main between 1 Sep and 2 Oct 2026 UTC.

Audited it by using it and caught what the coding agents missed: an age rule built so admitted rows could vanish before I had seen them, rebuilt the same day; and an aggregator endpoint rejected on a test of the wrong parameter, whose re-measurement, prompted by my question, took the feed from 93% irrelevant to every relevant posting, 420 of 420 stored. I approved a title-normalisation change only on a before-and-after comparison of every verdict, and of 1,458 titles exactly one changed, the one intended. I judged that postings reappearing under years-old dates were harvesting CVs, and the measurement afterwards found that 60 of the 83 Greenhouse postings already a week old at first sight were one role, reposted city by city since 2024. When an audit found a deletion path that would have erased my record of roles applied to, I set aside the proposed fix and designed my own.

Covered by 754 tests and 410 recorded mutations that deliberately break the code to prove the tests notice, with 10,381 lines of tests against 6,636 lines of source; the suite fails if a code change leaves any recorded mutation unable to run, and architectural rules are guarded by fitness functions in the ordinary suite. A production save failure on 27 Sep 2026 was found in the morning run log, reproduced, fixed and pushed before the next run, with the postings it missed recovered by the next walk.

Derived an original volatility baseline by log-log regression on 2020 to 2024 market data, ln(Vol%) ≈ 1.2 × ln(MarketCap) − 24.8, normalising each asset against its market-cap tier so genuine outliers surface instead of size artefacts.

Solved platform constraints: chunked storage across a 500KB per-key limit, execution-timeout safety with progress persistence and auto-resume, exchange symbol-prefix normalisation (1000SHIB to SHIB with price scaled), and time-weighted volatility correcting for partial-candle distortion.