Maneesh Ujji

Three confidence tiers, not two. Most systems do "confident" vs "not confident." I added a middle tier: the system still drafts, but explicitly flags its uncertainty so the reviewer knows to look closer.

Citations aren't optional. Every draft includes the source docs it pulled from. If the retrieval was weak or irrelevant, that's visible, so the reviewer can see the system's reasoning and not just its output.

Below 0.4, no draft at all. The system doesn't guess. It escalates directly to a human agent with the raw ticket. Some problems shouldn't have an AI answer.

Four specialized agents, not one general one. Each agent has a single job. The classifier doesn't draft. The drafter doesn't classify. This makes each component testable and replaceable independently.

Spam gets killed early. The classifier's first job is to stop spam from consuming any downstream compute. If it's spam, archive it and move on. No retrieval, no drafting, no logging.

Evaluation isn't afterthought. I built a full evaluation pipeline: accuracy per category, F1 scores, and confusion matrices against a baseline. If the classifier degrades, you know exactly where and by how much.

JSON memory, not a database. For a personal tool tracking a handful of routes, a lightweight JSON file beats setting up a database. The memory stores price snapshots over time: simple, portable, and zero infrastructure.

LLM for analysis, not just comparison. A simple "is today cheaper than yesterday?" check is trivial. I use Gemini to analyze trends across multiple snapshots: seasonality patterns, rate of change, and whether a drop is noise or a real deal.

Fully automated. Windows Task Scheduler triggers the agent daily. It fetches, analyzes, and emails, with no manual intervention. I wake up and either get a recommendation or nothing (no news = no action needed).

In the browser, not a separate app. Job seekers have 20 tabs open. Making them copy-paste into a separate tool breaks the flow. JobPulse runs right on the job listing page, with zero context switching.

Structured extraction, not raw text comparison. The LLM doesn't just compare blobs of text. It extracts structured fields (required skills, years of experience, tech stack) from the listing, then matches each against your resume's parsed skills.

Gaps matter more than matches. The most useful output isn't "you match 78%." It's "you're missing Kubernetes and they asked for it twice." That tells you whether to apply or what to study.

Four models, not one. The point wasn't to pick the best model; it was to show stakeholders that different approaches converge on similar accuracy, which builds confidence in the prediction itself.

Dashboard over notebook. The audience was non-technical. A Jupyter notebook is useless to them. The Plotly Dash dashboard lets them filter by launch site, orbit type, and booster version and see predictions update live.

SQL before modeling. Running SQL queries on the raw data first surfaced the patterns that mattered: launch site and booster version were the strongest predictors. The models confirmed what the data already showed.

Multi-category, not binary sentiment. "Positive vs negative" is a solved problem. Distinguishing joy from love, or anger from fear, in short informal text is where the real challenge is, and where NLP gets interesting.

Pipeline quality matters more than model choice. The preprocessing step (handling slang, emojis, hashtags, abbreviations) had more impact on accuracy than switching between models. Garbage in, garbage out.

Per-category evaluation, not just aggregate accuracy. 85% overall accuracy hides that the model might be terrible at detecting surprise (rare class) while great at joy (common class). I evaluated each emotion independently.

Clinical scoring, not arbitrary labels. PHQ-9 has established clinical cutoffs. I used the validated severity bands (minimal/mild/moderate/moderately severe/severe) rather than inventing my own thresholds.

Age as a lens, not a feature. Instead of just throwing age into the model, I used it to segment the analysis, revealing that severity patterns differ meaningfully across age groups, which has clinical implications.

Visualization for awareness, not diagnosis. This is a research tool, not a clinical one. The outputs are designed to communicate patterns to researchers and public health audiences, not to diagnose individuals.

Research question first, not technique first. I didn't start with "let me try regression." I started with "does diabetes prevalence predict COVID outcomes?" The method follows the question.

Confounders acknowledged, not ignored. The correlation between diabetes and COVID mortality is real, but income, obesity, and healthcare access travel with it. The analysis flags this instead of overstating the finding.

Audience-aware visualization. The outputs were designed for people who don't read scatter plots. Clear labels, annotated axes, and takeaway-first summaries rather than raw statistical output.