CV Review (Strict HR/Recruitment Screening)
Verdict
Current draft shows strong technical depth but reads as research notes instead of hiring-ready achievement evidence. Most bullets describe activity, not business or model impact. Immediate upgrade needed: stronger ownership verbs, hard metrics, and tighter Harvard-style scan structure.
1) Weak/Passive Verbs Replaced
- Present technical research progress -> Drove weekly research updates that influenced architecture decisions
- Collaborate with faculty -> Co-designed experiment plans with faculty and graduate researchers
- Maintain a dedicated research commitment -> Sustained 20+ weekly research hours while preserving top academic performance
- Engineered an end-to-end pipeline (good start) -> Architected a reproducible experimentation pipeline with tracked outcomes
- Integrated innovations and developed custom modules -> Implemented feature-decoupling modules that improved measurable generation quality
- Designed and implemented a novel loss function -> Introduced an identity-mapping loss to improve content consistency
- Streamlined collaboration -> Standardized dataset/checkpoint publishing to accelerate cross-team reuse
- Benchmarked performance -> Executed benchmark suites with controlled protocols and tracked variance
- Conducted comparative analysis -> Quantified looped-vs-standard LM trade-offs with perplexity and reasoning metrics
- Developed an RL agent -> Built and optimized an RL policy system under non-stationary demand
- Evaluated efficacy -> Validated SAC/PPO performance against baseline with callback-driven evaluation
2) Missing Metrics (What to Estimate or Retrieve)
For each project, fill at least 2-3 measurable outcomes:
- Model Quality:
- FID / LPIPS / SSIM deltas for generation tasks
- Top-1 accuracy / exact match / pass@k for reasoning tasks
- Reward delta vs random baseline for RL
- Efficiency:
- Training time reduction (hours or %)
- Throughput increase (steps/sec, samples/sec)
- GPU memory reduction (GB or %)
- Reliability/Reproducibility:
- Successful rerun rate (%)
- Failed run reduction (%)
- Checkpoint recovery success rate (%)
- Delivery Velocity:
- Experiment cycles per week
- Debugging time saved (hours/week)
- Time-to-result reduction (%)
- Reach/Adoption (Portfolio):
- Monthly visitors
- Search-to-click conversion
- Number of technical posts and average read time
3) Harvard CV Compliance Check
Status: Partially compliant before rewrite; compliant after rewrite in new file
What was fixed:
- Removed weak activity-heavy phrasing and switched to outcome-first bullets
- Enforced concise, high-signal sections and cleaner hierarchy
- Kept pronouns out (no I / me / my)
- Standardized role entries with institution, location, and dates
- Increased ATS scannability with explicit technical stacks and measurable placeholders
- Corrected contact line to show real email label instead of ambiguous wording
Remaining requirement for full excellence:
- Replace all [X] placeholders with verified numbers before external submission
4) Top 3 Bullets Rewritten with Golden Formula
- Achieved [X%] improvement in style-transfer fidelity as measured by [FID/LPIPS or human preference] by integrating Skeleton Distance Transform and DFT-based content-style decoupling modules.
- Achieved [X] benchmark runs/week as measured by automated evaluation throughput by building a reproducible lm-evaluation-harness pipeline with strict extraction controls.
- Achieved [X%] training speedup as measured by environment steps/second by engineering a custom Gymnasium simulator and accelerating bottlenecks with Numba.
Final Summary (Hiring Readiness)
- Strengths: clear AI/ML project depth, credible research experience, strong academic profile.
- Risks: insufficient quantified outcomes and occasional wording that signals participation rather than ownership.
- Result after rewrite: CV now reads as impact-oriented, ATS-friendly, and recruiter-scannable for AI/ML internship or entry-level research engineering roles.
- Final action before sending: replace all metric placeholders with real values from logs, leaderboards, or experiment trackers.