Choose a role
A role maps to supported technical domains. A job description can add relevance hints.
ADAPTIVE TECHNICAL INTERVIEW SYSTEM SMART
SMART
MOCK
INTERVIEW
Can technical interview answers be evaluated more consistently?
A local practice system that asks technical questions, compares answers with references, and turns each session into clear feedback.
Explore the system Candidate response Semantic comparison Reference prompt
LOCAL-FIRST · OPEN SOURCE01 08
01
THE PRODUCT
A practice session with visible limits.
Choose a role, work through technical questions, and review a record of what you answered. The system makes practice more structured; it does not decide whether someone should be hired.
Practice aidSimilarity to a reference, not a correctness grade.Local-firstAnswers and session results stay on the candidate’s computer.Evidence-ledExperimental checks are separately labelled and opt-in.
02
QUESTION BANK · READ-ONLY CORPUS
A broad bank, organized for practice.
The application loader verifies a curated base and four approved expansion batches before exposing questions to the interview flow.
LOADED KNOWLEDGE BASE666questions
Across eight technical domains and three difficulty levels.
8 domains3 levelsread-only
Machine Learning174
Generative AI & RAG94
Deep Learning86
MLOps & Model Deployment85
Data Analysis75
SQL54
NLP53
Python45
DIFFICULTY MIXFixed progression.
Fixed progression.
Clear expectations.
Mixed interviews follow Easy → Medium → Hard. That core ladder does not change in response to similarity scores.
Easy
125
Medium
435
Hard
106
CURATION NOTE The active loader combines 487 frozen v5 questions with 179 approved expansion records. The repository states that no question has been reviewed by a human subject-matter expert. The separate legacy /api system retains a 4,085-question bank; it is not the corpus used by the new /demo flow.
03
DATA → SESSION
From curated records to a usable interview.
The same validated corpus powers question selection, reference-answer comparison, and the session summary.
01Curated recordsQuestion · answer · concepts→
02Validate & fingerprintSchema · manifests · integrity→
03Select a sessionRole · domain · difficulty→
04Capture the responseLocal session store→
05Review the evidenceSimilarity · observations · report
src/retrieval/corpus.pysrc/demo/catalog.pysrc/demo/engine.pysrc/demo/store.py
04
INTERVIEW ENGINE
A guided conversation, not a quiz dump.
The new demo plans a structured technical interview, stores each turn locally, and builds the review from the questions and answers in that session.
↗
Get the next prompt
Curated questions are selected for the role, topic, and planned difficulty.
Answer in context
Write prose or use the code editor for the practical exercise. Save and return later.
Review separate signals
Sentence-BERT similarity, exact-word observations, and optional experiments remain distinct.
Continue and reflect
Answer-aware follow-ups may probe missing concepts; the core difficulty ladder stays fixed.
Fixed Easy → Medium → Hard
Difficulty is planned in advance. Similarity and optional evaluators never control the next question’s difficulty.
Easy→Medium→Hard
Responsive prompts, bounded scope
Default follow-ups use curated concepts and the candidate’s answer. A local-model follow-up mode is optional, explicitly experimental, and not human-validated.
OPTIONAL · EXPERIMENTALLEGACY /api ROUTE The original interview flow is separate and retains adaptive rounds driven by similarity scores. It uses the legacy 4,085-question bank.
05
ANSWER SCORING
One active scorer. Two legacy experiments.
The new interview demo uses Sentence-BERT only. Older scorers remain in the repository’s legacy model registry, but they are not selectable for new demo sessions.
01 / ACTIVE IN NEW DEMODEFAULT
Sentence-BERT
all-MiniLM-L6-v2
Embeds the candidate answer and reference answer, then compares their semantic similarity. Reworded answers may resemble the reference, but a high similarity is still not proof of correctness.
02 / LEGACY REGISTRYNOT IN NEW DEMO
Siamese Bi-LSTM
Custom PyTorch answer-similarity model trained on the project’s question/answer pairs. The registry notes it mainly saw exact-match positives, so it can be stricter with reworded responses.
Answer pair→Twin scorer→Similarity
03 / LEGACY REGISTRYNOT IN NEW DEMO
TF-IDF + grammar
Classic text-overlap similarity with a grammar check. The registry’s comparable score is the TF-IDF similarity; grammar is not part of that number.
Word overlap+Grammar check→Baseline
REUSABLE LEGACY INTERFACEAnswer pair → model registry → selected scorer → similarity output
All model scores measure resemblance to a reference response. None have been validated against expert human marking.
06
FEEDBACK & REPORTING
Show the candidate what the system can actually see.
A session summary brings together the transcript and modest evidence signals without turning them into a hiring recommendation.
ANSWERSimilarity detail↓Terms found / not found as written↓Topics + practice areas↓SESSION SUMMARY + JSON EXPORT
Clarity over false certainty.
Similarity is labelled as similarity. Key-concept checks report wording found or not found; absence does not mean an idea was absent, and presence does not prove understanding.
The report records questions, answers, topics, observations, practice areas, and a turn-by-turn transcript. No overall grade is produced.
MAIN DEMOLocal report · JSON exportLEGACY APIPDF and email delivery modules
07
REAL PROJECT UI
Screens from the working project.
These are repository screenshots of the real Flask demo, not generated interface mockups.
08
ARCHITECTURE & TOOLING
A local application, with experiments around it.
The interview path, opt-in research surfaces, and developer tooling are kept distinct in the implementation and in this case study.
01CandidateBrowser · local session
↓02Flask demoRole · interview plan · persistence
↓03Validated question corpus666 records · read-only · fingerprinted
↓04A · DEFAULTSentence-BERTReference similarity
04B · OPT-INModular evaluation / code checkExperimental, separate output
04C · SEPARATERetrieval diagnosticsMiniLM + FAISS / BM25; not answer grading
05Session reportObservations · practice areas · JSON export
VERIFIED TECHNOLOGY
Tooling labels show where each technology belongs.
PythonApplicationFlaskApplicationSentence TransformersScoringall-MiniLM-L6-v2ScoringFAISS + BM25Opt-in retrievalPyTorchLegacy scorerscikit-learnLegacy scorerPandas · NumPyDataHTML · CSS · JavaScriptInterfaceDocker · MLflow · DVCDeveloper toolingGitHub ActionsCI
FUTURE DEVELOPMENT / VALIDATION
Human subject-matter review of the question bank, human-rated answer evaluation, and usability research remain needed before any score can be treated as an assessment.
Structured practice.
Honest signals.
Smart Mock Interview combines a curated technical question bank, semantic answer comparison, answer-aware follow-ups, and session reporting in one local practice system.
The current demo’s similarity score is not validated against human marking. Experimental evaluation and retrieval are separate from the default scoring path.
RESEARCH PROTOTYPE · LOCAL-FIRST · NO HIRING DECISIONS© 2026 MOHAN RAMAMURTHY